ABSTRACT. In this talk we present a family of models for the simulation of earthquake, or landslides generated tsunamis. All of them fit in the framework of shallow-flows. Here, the flow is supposed to be modeled by shallow-water type systems composed by one or several layers of fluid. Multilayer shallow-water models allow us to recover the vertical profile of the velocities, that may be relevant at the early stages of the landslide-fluid interaction, as well as non-hydrostatic corrections. Earthquake generated tsunamis are supposed to be driven either by a simple Okada model or for more sophisticated ones, like SeisSol model. Concerning the evolution of the landslide, either it is considered to be a rigid body and its motion it is supposed to be known, either it is supposed to be modeled by a Savage-Hutter type model. The resulting models are discretized using a high-order finite-volume path-conservative scheme and implemented on a multi-GPU framework. Finally, an exhaustive validation procedure has been carrying out by the comparison with laboratory experiments and real events over real bathymetries.

ABSTRACT. We have been using various environments and spaces to meet high transportation demands. However, traffic congestion, deteriorating transportation infrastructure, and environmental pollution have become current social problems. To solve these problems, flying vehicles that use near ground space (NGS) are attracting attention. In order to develop such vehicles efficiently, highly accurate computer simulation technology is required. In this study, computer simulations are performed by coupling fluid and rigid body motions using two calculation methods. One is the moving computational domain method, in which the object and computational domain are moved as a single unit to represent the motion of the object and the flow around the body, and the other is the multi-axis sliding mesh method, in which physical quantities are transferred at the boundaries to reproduce the motion of objects with different motions, such as rotating parts. Because the flying car in the development stage is small and has a shape that obtains thrust from multiple propellers, the insertion of disturbances was considered because of the possible effects of wind on the aircraft during actual flight. In this study, we attempted to clarify the effect of wind on the flying car by performing flight simulations in six patterns, one with no wind and the other with a disturbance inserted that causes a headwind during forward flight and a cross-wind during turning flight.

ABSTRACT. While the unstable turbulence condition known as Vortex Ring State (VRS) in rotorcraft has been studied mainly in helicopters, there have not been many studies of drones, which have become more active in recent years. In particular, there are few studies using numerical simulations focusing on small diameter rotors such as quadcopters. In this paper, descent simulations are performed using a rotor model with a diameter of 8 inches and a propeller pitch of 4.5 inches, which is used for quadcopters. In this study, Moving Computational Domain (MCD) method is used to reproduce the descent motion of the rotor over the entire moving computational domain. In addition, Sliding mesh method is applied to reproduce the rotor rotation within the computational grid. This method allows flow field analysis under free rigid body motion of the analytical model. The displacement of the computational domain itself is applied at each step to reproduce the descent conditions. By combining these methods, fluid flow simulations under vertical descent and conditions are performed to visualize the flow field. Flow field evaluation using the Q criterion showed that even in a small rotor, circulating vortices are generated at a velocity close to the induced velocity vh. VRS was also observed around the rotor under the conditions of horizontal speed VH=2.0vh, descent speed VY=1.0vh, and descent angle of 26.6 deg, but at the same time vortex divergence was also observed. It was inferred that the forward velocity component helps to avoid VRS.

ABSTRACT. Design of antennas for modern applications is a challenging task that combines cognition-driven development of topology intertwined with tuning of its parameters using rigorous numerical optimization. However, the process can be streamlined by neglecting the engineering insight in favor of automatic de-termination of structure geometry. In this work, a specification-oriented design of topologically agnostic antenna is considered. The radiator is developed using a bi-stage algorithm that involves min-max classification of randomly-generated topologies followed by local tuning of the promising designs using a trust-region optimization applied to a feature-based representation of the structure frequency response. The automatically generated antenna is characterized by –10 dB bandwidth of over 600 MHz w.r.t. the center frequency of 6.5 GHz and a dual-lobe radiation pattern. The obtained performance figures make the radiator of use for in-door positioning applications. The design method has been favorably compared against the frequency-based trust-region optimization.

ABSTRACT. Antenna far-field performance is normally evaluated in expensive laboratories that maintain strict control over the propagation environment. Alternatively, the responses can be measured in uncontrolled conditions and then refined to extract useful information on the structure field-related behavior. Here, a framework for correction of antenna measurements performed in non-anechoic test site has been proposed. The method involves automatic synchronization (in time-domain) of spatially separated measurements followed by their combination so as to augment the fraction of the signal that represents antenna performance while suppressing the unwanted interferences. The method has been demonstrated based on a total of six experiments performed in a standard office room and favorably compared against the state-of-the-art techniques from the literature.

ABSTRACT. This research introduces a detailed computational framework designed to analyze and forecast the complex multiphase characteristics of eco-friendly lead-free piezoelectric materials, which are essential for developing sustainable technological advancements. Lead-free piezoelectric materials have a significant thermo-electromechanical response, although their electromechanical characteristics vary throughout different phases of the material. Lead-free piezoelectric materials undergo phase changes, including rhombohedral (R3c), orthorhombic (Pnma), tetragonal (P4bm), and cubic (Cc) phases, when the temperature changes. These phases are determined by the symmetry and alignment of the ferroelectric domains. Furthermore, multiple phases exist simultaneously under certain temperature, electrical, and mechanical conditions, resulting in the material displaying intricate multiphase behavior. Studying such behaviour is crucial for evaluating the performance of these materials. The computational approach in this research relies on Landau-Ginzburg-Devonshire theory to model micro-domain phase changes in the material. This research will enhance our comprehension of the significance of complex multiphase behaviour in creating environment-friendly and sustainable technological solutions.

ABSTRACT. We address the research gap in evaluating the effectiveness of network seeding strategies in maximizing the spread of beliefs within non-progressive competing complex contagions. Our study focuses on management perspective of devising promotional campaigns for sustainable and health behaviors. By comparing recent and established methods for identifying central nodes in social networks our computational analysis, conducted on two empirical datasets shows that it is possible to achieve widespread adoption of socially responsible attitudes even under limited network information. However, this success requires a strategic management approach that includes additional efforts to prevent the targeted influencers from abandoning these attitudes in the future.

ABSTRACT. In response to the growing threat of Advanced Persistent Threat (APT) in network security, our research introduces an innovative APT malware attribution tool, the APTMalKG knowledge graph. This knowledge graph is constructed from comprehensive APT malware data and is refined through a multi-stage graph clustering process. To enhance its effectiveness, we have incorporated domain-specific meta-paths into the GraphSAGE graph embedding algorithm. Our approach includes an ontology model capturing complex APT malware characteristics and behaviors, extracted from sandbox analysis reports and expanded intelligence. To manage the graph's granularity and scale, we categorize nodes based on domain knowledge, form a correlation subgraph, and progressively adjust similarity thresholds and edge weights. The refined graph maintains crucial attribution data while reducing complexity. By integrating domain-specific meta-paths into GraphSAGE, we achieve improved APT attribution accuracy (average ACC 91.16%, average F1 score 89.82%, average AUC 98.99%) and enhanced performance (graph scale reduced by an order of magnitude). This study not only benefits network security analysts with an intuitive knowledge graph but also explores large-scale graph computing methods for practical scenarios, offering a multi-dimensional perspective on APT malware analysis and attribution research, highlighting the value of knowledge graphs in network security.

ABSTRACT. This paper presents a closed form solution to the eigen-problem of combinatorial graph Laplacian for a new type of regular grid graphs - biweighted grid graphs. Biweighted grid graphs differ from ordinary ones in that the weights along a single dimension are altering which adds complexity to the eigen-solutions and makes the graphs better test-bed for potential applications.

ABSTRACT. The paper presents a simulation model allowing examination of cooperation between two telecommunication networks concepts: IP Multimedia Subsystem/Next Generation Network (IMS/NGN) and Software-Defined Networking (SDN). Application of the SDN architecture elements in currently used IMS/NGN networks unifies control and management of transport resources for various transport technologies and equipment manufacturers. The structure of the modeled multidomain network and details about the simulator operation are described. Tests proving correctness of its operation are carried out. Selected research results regarding mean Call Set-up Delay and mean Call Disengagement Delay in the considered network are presented demonstrating that the cooperation between IMS/NGN and SDN is possible.

ABSTRACT. With the depletion of IPv4 address resources, the prevalence of IPv4 address leasing services by hosting providers has surged. These services allow users to rent IP blocks, offering an affordable and flexible solution compared to traditional IP address allocation. Unfortunately, this convenience has led to an increase in abuse, with illegal users renting IP blocks to host malicious content such as phishing sites and spam services. To mitigate the issue of IP abuse, some research focuses on individual IP identification for point-wise blacklisting. However, this approach leads to a game of whack-a-mole, where blacklisted IPs become transient due to content migration within the IP block. Other studies take a block perspective, recognizing and classifying IP blocks. This enables the discovery of potentially malicious IPs within the block, effectively countering service migration issues. However, existing IP block identification methods face challenges as they rely on specific WHOIS fields, which are sometimes not updated in real-time, leading to inaccuracies. In terms of classification, methods rely on limited statistical features, overlooking vital relationships between IP blocks, making them susceptible to evasion. To address these challenges, we propose BlockFinder, a two-stage framework. The first stage leverages the temporal and spatial stability of services to identify blocks of varying sizes. In the second stage, we introduce an innovative IP block classification model that integrates global node and local subgraph representations to comprehensively learn the graph structure, thereby enhancing evasion difficulty. Experimental results show that our approach achieves state-of-the-art performance.

ABSTRACT. Water movement in soil is essential for weather monitoring, prediction of natural disasters, and agricultural water management. Richardson-Richards' equation (RRE) is the characteristic partial differential equation for studying soil water movement. RRE is a non-linear PDE involving water potential, hydraulic conductivity, and volumetric water content. This equation has underlying non-linear parametric relationships called water retention curves (WRCs) and hydraulic conductivity functions (HCFs). This two-level non-linearity makes the problem of unsaturated water flow of soils challenging to solve. Physics-Informed Neural Networks (PINNs) offer a powerful paradigm to combine physics in data-driven techniques. From noisy or sparse observations of one variable (water potential), we use PINNs to learn the complete system, estimate the parameters of the underlying model, and further facilitate the prediction of infiltration and discharge. We employ training on RRE, WRC, HCF, and measured values to resolve two-level non-linearity directly instead of explicitly deriving water potential or volumetric water content-based formulations. The parameters to be estimated are made trainable with initialized values. We take water potential data from simulations and use this data to solve the inverse problem with PINN and compare estimated parameters, volumetric water content, and hydraulic conductivity with actual values. We chose different types of parametric relationships and wetting conditions to show the approach's effectiveness.

ABSTRACT. We present a new deep learning paradigm for the generation of sparse approximate inverse (SPAI) preconditioners for matrix systems arising from the mesh-based discretization of elliptic differential operators. Our approach is based upon the observation that matrices generated in this manner are not arbitrary, but inherit properties from differential operators that they discretize. Consequently, we seek to represent a learnable distribution of high-performance preconditioners from a low-dimensional subspace through a carefully-designed autoencoder, which is able to generate SPAI preconditioners for these systems. The concept has been implemented on a variety of finite element discretizations of second- and fourth-order elliptic partial differential equations with highly promising results.

ABSTRACT. We propose a novel multi-agent system approach to solve a coverage problem in Wireless Sensor Networks (WSN) based on the collective behavior of (ϵ,h)-Learning Automata (LA). The coverage problem can be stated as a request to find a minimal number of sensors spending energy of their batteries to provide the requested level of coverage of the whole monitored area. We propose a distributed self-organizing algorithm based on the participation of LA in an iterated Spatial Prisoner’s Dilemma game. We show that agents achieve a solution corresponding to Nash equilibrium, which provides maximization of not known for agents a global criterion related to the requested level of the coverage with a minimal number of sensors which turn on their batteries.

ABSTRACT. Chinese spelling correction (CSC) is an important task in the field of natural language processing (NLP). While existing state- of-the-art methods primarily leverage pre-trained language models and incorporate external knowledge sources such as confusion sets, they often fall short in fully leveraging local information that surrounds erroneous words. In our research, we aim to bridge a crucial gap by introducing a novel CSC model that is enhanced with a Gaussian attention mecha- nism. This integration allows the model to adeptly grasp and utilize both contextual and local information. The model incorporates a Gaussian at- tention mechanism, which results in attention weights around erroneous words following a Gaussian distribution. This enables the model to place more emphasis on the information from neighboring words. Addition- ally, the attention weights are dynamically adjusted using learnable hy- perparameters, allowing the model to adaptively allocate attention to different parts of the input sequence. In the end, we adopt a homophonic substitution masking strategy and fine-tune the BERT model on a large- scale CSC corpus. Experimental results show that our proposed method achieve a new state-of-the-art performance on the SIGHAN benchmarks.

ABSTRACT. Designing microwave components involves managing multiple objectives such as center frequencies, impedance matching, and size reduction for miniaturized structures. Traditional multi-objective optimization (MO) approaches heavily rely on computationally expensive population-based methods, especially when executed with full-wave electromagnetic (EM) analysis to guarantee reliability. This paper introduces a novel and cost-effective MO technique for microwave passive components utilizing a machine learning (ML) framework with artificial neural network (ANN) surrogates as the primary prediction tool. In this approach, multiple candidate solutions are extracted from the Pareto set via optimization using a multi-objective evolutionary algorithm (MOEA) applied to the current ANN model. These solutions expand the dataset of available (EM-simulated) parameter vectors and refine the surrogate model iteratively. To enhance computational efficiency, we employ variable-resolution EM models. Tested on two microstrip circuits, our methodology competes effectively against various surrogate-assisted methods. The average computational cost of the algorithm is below three hundred high-fidelity EM analyses of the circuit, while the quality of generated Pareto sets surpasses those produced by the benchmark methods.

ABSTRACT. The design of antenna systems poses a significant challenge due to stringent performance requirements dictated by contemporary applications and the high computational costs associated with models, particularly full-wave electromagnetic (EM) analysis. Presently, EM simulation plays a crucial role in all design phases, encompassing topology development, parametric studies, and the final adjustment of antenna dimensions. The latter stage is especially critical as rigorous numerical optimization becomes essential for achieving optimal performance. In an increasing number of instances, global parameter tuning is necessary. Unfortunately, the use of nature-inspired algorithms, the prevalent choice for global design, is hindered by their poor computational efficiency. This article presents an innovative approach to cost-efficient global optimization of antenna input characteristics. Our methodology leverages response feature technology, ensuring inherent regularization of the optimization task by exploring the nearly-linear dependence between the coordinates of feature points and the antenna's dimensions. The optimization process is structured as a machine learning (ML) procedure, utilizing a kriging surrogate model rendering response features to generate promising candidate de-signs (infill points). This model is iteratively refined using accumulated EM simulation data. Further acceleration is achieved by incorporating multi-fidelity EM analysis, where initial sampling and surrogate model construction use low-fidelity EM simulations, and the ML optimization loop employs high-fidelity EM analysis. The multi-fidelity EM simulation data is blended into a single surrogate using co-kriging. Extensive verification of the presented algorithm demonstrates its remarkable computational efficiency, with an average running cost not exceeding ninety EM simulations per run and up to a seventy percent relative speedup over the single-fidelity procedure.

ABSTRACT. Complete Coverage and Path Planning methods operate on many models depending on initial constraints and user demands. In this case, we optimize paths for a set of UAVs in the disaster area divided into rectangular regions of different sizes and priorities representing the expected number of victims. Paths maximize the number of victims localized in the first minutes of the UAVs' operation and minimize the entire operation makespan. The problem belongs to the domain of multiobjective optimization; therefore, we apply the Strength Pareto Evolutionary Algorithm 2, which is equipped with several problem-specific perturbation operators. In the experimental part, we use SPEA2 to four selected test cases from a TCG-CCPP generator powered by actual data on residents in selected regions in Poland published by Statistics Poland.

ABSTRACT. Given an optimization problem, combining knowledge from both (i) structural or algorithmic known results and (ii) new solving techniques, helps gain insight and knowledge on the aforementioned problem by tightening the gap between lower and upper bounds on the optimal value. Additionally, this gain may be further improved by iterating (i) and (ii) until a fixed point is reached. In this paper, we illustrate the above through the classical Cyclic Bandwidth problem, an optimization problem which takes as input an undirected graph G = (V,E) with |V | = n, and asks for a labeling φ of V in which every vertex v takes a unique value φ(v) ∈ [1; n], in such a way that Bc (G, φ) = max{minuv∈E(G) {|φ(u)−φ(v)|, n−|φ(u)−φ(v)|}}, called the cyclic bandwidth of G, is minimized. Using the classic benchmark from the the Harwell-Boeing sparse matrix collection introduced in [16], we show how to combine (a) previous results from the Cyclic Bandwidth literature, and (b) new solving techniques which we first present, and then implement starting from the best results obtained in step (a). We show that this (possibly iterated) process improves the best known bounds for a large number of instances from our benchmark, and actually allows us to determine the optimal cyclic bandwidth value for half of these instances.

ABSTRACT. The significance of learning constraints from data is underscored by its potential applications in real-world problem-solving. While constraints are popular for modeling and solving, the approaches to learning constraints from data remain relatively scarce. Furthermore, the intricate task of modeling demands expertise and is prone to errors, thus constraint acquisition methods offer a solution by automating this process through learnt constraints from examples or behaviours of solutions and non-solutions. This work introduces a novel approach grounded in Deep Neural Network (DNN) based on Symbolic Regression that, by setting suitable loss functions, constraints can be extracted directly from datasets. Using the present approach, direct formulation of constraints was achieved. Furthermore, given the broad pre-developed architectures and functionalities of DNN, connections and extensions with other frameworks could be foreseen.

ABSTRACT. In this research, we investigate the effectiveness of Large Language Models (LLMs) in answering health-related questions. The rapid growth and adoption of LLMs, such as ChatGPT, have raised concerns about their accuracy and robustness in critical domains such as Health Care and Medicine. We conduct a comprehensive study comparing multiple LLMs, including recent models like GPT-4 or Llama2, on a range of binary health-related questions. Our evaluation considers various context and prompt conditions, with the objective of determining the impact of these factors on the quality of the responses. Additionally, we explore the effect of in-context examples in the performance of top models. To further validate the obtained results, we also conduct contamination experiments that estimate the possibility that the models have ingested the benchmarks during their massive training process. Finally, we also analyse the main classes of errors made by these models when prompted with health questions. Our findings contribute to understanding the capabilities and limitations of LLMs for health information seeking.

ABSTRACT. Machine learning on large-scale genomic or transcriptomic data is important for many novel health applications. For example, precision medicine tailors medical treatments to patients on the basis of individual biomarkers, cellular and molecular states, etc. However, the data required is sensitive, voluminous, heterogeneous, and typically distributed across locations where dedicated machine learning hardware is not available. Due to privacy and regulatory reasons, it is also problematic to aggregate all data at a trusted third party. Federated learning is a promising solution to this dilemma, because it enables decentralized, collaborative machine learning without exchanging raw data. In this paper, we perform comparative experiments with the federated learning frameworks TensorFlow Federated and Flower. Our test case is the training of disease prognosis and cell type classification models. We train the models with distributed transcriptomic data, considering both data heterogeneity and architectural heterogeneity. We measure model quality, robustness against privacy-enhancing noise, computational performance and resource overhead. Each of the federated learning frameworks has different strengths. However, our experiments confirm that both frameworks can readily build models on transcriptomic data, without transferring personal raw data to a third party with abundant computational resources.

ABSTRACT. The development of black-box survival models has created a need for methods that explain their outputs, just as in the case of traditional machine learning methods. Survival models usually predict functions rather than point estimates. This special nature of their output makes it more difficult to explain their operation. We propose a method to generate plausible counterfactual explanations for survival models. The method supports two options that handle the special nature of survival models' output. One option relies on the Survival Scores, which are based on the area under the survival function, which is more suitable for proportional hazard models. The other one relies on Survival Patterns in the predictions of the survival model, which represent groups that are significantly different from the survival perspective. This guarantees an intuitive well-defined change from one risk group (Survival Pattern) to another and can handle more realistic cases where the proportional hazard assumption does not hold. The method uses a Particle Swarm Optimization algorithm to optimize a loss function to achieve four objectives: the desired change in the target, proximity to the explained example, likelihood, and the actionability of the counterfactual example. Two predictive maintenance datasets and one medical dataset are used to illustrate the results in different settings. The results show that our method produces plausible counterfactuals, which increase the understanding of black-box survival models.

ABSTRACT. Decision support systems in the medical domain is budding field that aims to improve healthcare and overall recovery for patients. While treatment remains specific to individual symptoms, the diagnosis of patients is fairly general. Interpreting the diagnosis and assigning the appropriate care treatment is a crucial part undertaken by medical professionals, however, in critical scenarios, having access to recommendations from a clinical decision support system may prove life-saving. We present a real-world application of inverse reinforcement learning (IRL) to assess the implicit cognitive state of doctors when evaluating decision support data on a patient's risk of acquiring Type 2 Diabetes mellitus (T2DM). We show the underlying process of modelling a Markov Decision Process (MDP) using real-world clinical data and experiment with various policies extracted from sampled trajectories. The results provide insight to approach modelling real-world data into interpretable solutions via IRL.

ABSTRACT. This paper approaches a new model for arrhythmia diagnosis based on short-duration electrocardiogram (ECG) heartbeats. To detect 8 arrhythmia classes efficiently, we design a Deep Learning model based on the Focal modulation layer. Moreover, we develop a distance variation of the SMOTE technique to address the problem of data imbalance. The classification algorithm includes a block of Residual Network for feature extraction and an LSTM network with a Focal block for the final class prediction. The approach is based on the analysis of variable-length heartbeats from leads MLII and V5, extracted from 48 records of the MIT-BIH Arrhythmia Database. The methodology’s novelty consists of using the focal layer for ECG classification and data augmentation with DTW distance (Dynamic Time Warping) using the SMOTE technique. The presented approach offers real-time classification and is simple since it combines feature extraction, selection, and classification in one stage. Using data augmentation with SMOTE variant and Focal-based Deep learning architecture to identify 8 types of heartbeats, the method achieved an impressive overall accuracy, F1-score, precision, and recall of 98.61%, 94.08%, 94.53%, and 93.68% respectively. Additionally, the classification time per sample was only 0.002 seconds. Therefore, the suggested approach can serve as an additional tool to aid clinicians in ensuring rapid and real-time diagnosis for all patients with no exclusivities. Thus, helping to reduce healthcare disparities by making diagnostic tools more accessible and effective for a wide range of communities. This contributes to a fairer healthcare system and a more sustainable society.

ABSTRACT. Accurate retrodictions of the past evolution of convection in the Earth’s mantle are crucial to obtain a qualitative understanding of this central mechanism behind impactful geological events on our planet. They require highly resolved simulations and therefore extremely scalable numerical methods. This paper applies the massively parallel matrix-free finite element framework HyTeG to approximate solutions to stationary Stokes systems and time-dependent, coupled convection problems. It summarizes the underlying mathematical model and verifies the implementation through semi-analytical setups and community benchmarks. The numerical results agree with the expected outcomes from the literature.

ABSTRACT. There is a strong need for faster explicit elastic wavefield simulations for large and complex three-dimensional media using a structured finite element method. Such wavefield simulations are suitable for GPUs, which have been improving their computational performance in recent years, and the use of GPUs is expected to speed up such simulations. However, there is still room for speeding up such simulations using GPUs, since the performance of GPUs is not fully exploited just by its simple use, and the conventional method involves some numerical dispersion. In this paper, we propose a method for fast explicit structured-mesh wavefield simulation on GPUs by utilizing int8-Tensor Cores and reducing numerical dispersion. The proposed method was implemented for GPUs, and its performance was evaluated on an application example that simulates a real-world problem, showing that it is 17.0 times faster than the conventional method.

ABSTRACT. The Joint Center for Satellite Data Assimilation (JCSDA) is a multi-agency research center established to improve the quantitative use of satellite data in atmosphere, ocean, climate and environmental analysis and prediction systems. At the JCSDA, scientists and software engineers within the Joint Effort for Data Assimilation Integration (JEDI) are developing a unified data assimilation framework for research and operational use. To harness the full potential of ever-increasing volumes of data from new and evolving Earth observation systems, a new online database software service has been developed and deployed by the JCSDA. As a core component of JEDI, the Research Repository for Data and Diagnostics (R2D2) performs data registration, management, and configuration services for data assimilation computational workflows. We present an overview of R2D2’s distributed system of data stores, SQL data model, intuitive python API, and user support efforts. In addition, we will detail R2D2’s utilization by environmental prediction applications developed by the JCSDA and its partners.

ABSTRACT. The ground structure can substantially influence seismic ground motion underscoring the need to develop a ground structure model with sufficient reliability in terms of ground motion estimation for earthquake damage mitigation. While many methods for generating ground structure models have been proposed and used in practice, there remains room for enhancing their reliability. In this study, amid many candidate 3D ground structure models generated from geotechnical engineering knowledge, we propose a method for selecting a credible 3D ground structure model capable of reproducing observed earthquake ground motion, utilizing seismic ground motion data solely observed at the ground surface and employing 3D seismic ground motion analysis. Through a numerical experiment, we illustrate the efficacy of this approach. By conducting 100-1000 cases of fast 3D seismic wave propagation analyses using graphic processing units (GPUs), we demonstrate that a credible 3D ground structure model is selected according to the quantity of seismic motion information. We show the effectiveness of the proposed method by comparing the accuracy of estimated seismic motions using ground structure models that were selected versus those that were not selected from the pool of candidate models.

ABSTRACT. In this article, we propose a technique for finding a solution for large unsymmetric linear systems. Such problems are related to different areas e.g., image processing, computer vision, and structural engineering. The classic Biconjugate Gradient Stabilized (BiCGStab) method was introduced as a rapid and smoothly converging alternative to both the BiCG method and the CGS method. Following the initial proposal, this method was further optimized, resulting in the pipelined BiCGStab variant. This refined approach enhances scalability on distributed memory machines, yielding to substantial speed improvements compared to the standard BiCGStab method. However, it’s worth noting that the pipelined BiCGStab algorithm sacrifices some accuracy, which is stabilized with the residual replacement technique. This paper aims to address this issue by employing the ExBLAS-based reproducible approach. We validate the idea on a set of matrices from the SuiteSparse Matrix Collection.

ABSTRACT. In recent years, unsupervised and self-supervised graph representation learning has gained popularity in the research community. However, most proposed methods are focused on homogeneous networks, whereas real-world graphs often contain multiple node and edge types. Multiplex graphs, a special type of heterogeneous graphs, possess richer information, provide better modeling capabilities and integrate more detailed data from potentially different sources. The diverse edge types in multiplex graphs provide more context and insights into the underlying processes of representation learning. In this paper, we tackle the problem of learning representations for nodes in multiplex networks in an unsupervised or self-supervised manner. To that end, we explore diverse information fusion schemes performed at different levels of the graph processing pipeline. The detailed analysis and experimental evaluation of various scenarios inspired us to propose improvements in how to construct GNN architectures that deal with multiplex graphs.

ABSTRACT. Accurate and robust assessment of non-traditional approach-es used for training students and professionals in improving laparoscopic surgical skills has been attracting many research studies recently. Such assessment is particularly critical with the recent advances related to virtual environments and AI tools in addressing the need to expand the education and training in the medical domains. Network models and population analysis methods have been identified as excellent approaches in providing the much-needed assessment. This study aims at further advancing the surgical skill assessment by introducing a comparative approach to threshold optimization in analyzing the network models. While the majority of network methods often on arbitrary or hard thresholds for network construction and analysis, this research explores the efficacy of network-based parameters for identifying key elements and clusters in extracting useful information from the constructed networks. We report the positive impact of using network structural parameters, such as edge betweenness and modularity, to conduct robust analysis of the assessment networks. In this work, we employ electromyography (EMG) data and the NASA Task Load Index (NASA-TLX) scores for comprehensive skill evaluation. We present a case study that highlights the advantage of selecting thresholds based on the highest edge betweenness associated with the obtained assessment networks. This proposed approach method proved to be more effective in identifying participants who exhibit significant learning progression, aligning their muscle activation patterns closely with top performers. We demonstrate that optimizing thresholds through edge betweenness offers a more accurate visualization and assessment of skill acquisition in laparoscopic surgery training.

ABSTRACT. Graph embeddings have emerged as a powerful tool for representing complex network structures in a low-dimensional space, enabling the use of efficient methods that employ the metric structure in the embedding space as a proxy for the topological structure of the data. In this paper, we explore several aspects that affect the quality of a vertex embedding of graph-structured data. To this effect, we first present a family of flexible distance functions that faithfully capture the topological distance between different vertices. Secondly, we analyze vertex embeddings as resulting from a~fitted transformation of the distance matrix, rather than as a~direct result of optimization. Finally, we evaluate the effectiveness of our proposed embedding constructions by performing community detection on a host of benchmark datasets. The reported results are competitive with classical algorithms that operate on the entire graph while benefitting from a substantially reduced computational complexity due to the reduced dimensionality of the representations.

ABSTRACT. The recent rapid advancements of high-throughput sequencing technologies have made it possible for researchers to explore the microbial universe in high degrees of depth that were not possible even few years ago. Microbial communities occupy numerous environments everywhere and significantly impact the health of the living organisms in their environments. With the availability of microbiome data, advanced computational tools are critical to conduct the needed analysis and allow researchers to extract meaningful knowledge leading to actionable decisions. However, despite many attempts to develop tools to analyze the heterogeneous datasets associated with various microbiomes, such attempts lack the sophistication and robustness needed to efficiently analyze these complex heterogeneous datasets and produce accurate results. In addition, almost all current methods employ heuristic concepts that do not guarantee the robustness and reproducibility needed to provide the biomedical community with trusted analysis that lead to precise data-driven decisions. In this study, we present a network model that attempts to overcome these challenges by utilizing graph-theoretic concepts and employing multiple computational methods with the goal of conducting robust analysis and produce accurate results. To test the proposed model, we performed the analysis on plant microbiome datasets to obtain distinctive functional modules based on key microbial interrelationships in a given host environment. Our findings establish a framework for a new understanding of the association between functional modules based on microbial community structure.

ABSTRACT. In this paper we focused on a clustering task of fMRI data from a publicly available dataset of patients with mild depression. We analyzed these data with a popular tool, the CONN toolbox, to extract networks which later we aligned through a global multiple alignment software, MULTIMAGNA++ and then used this data to perform a clustering task. Our results show that is possible to achieve good results in correctly clustering patients from this data and possibly we can find another hidden group among control and affected by disease.

ABSTRACT. The formalism of multilayer networks (MLN) makes possible to model and understand the multiple relationships between entities in a system. Indeed, this representation has found its way into a wide range of disciplines, particularly in the fields of neuroscience and neuroimaging. Human brain modelling made possible the identification of the basis for the construction of morphological, structural and functional brain connectivity networks. In this work, we propose the design and implementation of a software pipeline for the construction and analysis of multilayer brain networks. This approach aims to identify groups of strongly connected nodes within the network and to evaluate the resulting communities. We examined 10 healthy subjects and 10 patients with multiple sclerosis. We analyzed the brain MLN by applying community detection algorithm that identified recurrent communities in patients with multiple sclerosis. To assess the structure of communities within the network, we calculate modularity indices for each subject. Finally, we confirm what has already been found in the literature, i.e. a high modularity in the brain networks of diseased subjects compared to those of healthy subjects. Future developments could involve aligning these networks to identify common patterns among multiple sclerosis patients and potentially identify subgroups of patients with similar neural characteristics.

ABSTRACT. Forward error correction is vital for communication since it allows error rate or required SNR reduction. With longer codes better correction ratio can be achieved Iterated codes are a solution for the construction of long code with a simple coder and decoder. Nevertheless, even though a basic iterative code decoder is simple, it cannot fully exploit the code’s potential, since some error patterns within the code’s correction capacity cannot be corrected. We propose two neural network assisted decoders: the first based on a classical neural network, and the second employing a convolutional neural network. The conducted experiments demonstrated that iterated codes can be decoded just by using such stand-alone neural network-based solutions. Additionally, we observed that even if the proposed neural decoders failed to correctly decode a code word, they were able to decrease the number of errors in the decoded message. Based on this observation we proposed an iterative neural network-based decoder. The resulting decoder showed significantly improved overall performance exceeding the performance of the classical decoder and proved that neural networks can be efficiently applied to iterative code decoding.

ABSTRACT. The paper considers the sigmoid activation function implementations for the artificial neural network hardware systems. The rational fraction number system is proposed to calculate this function. Such data presentation offers several benefits, including increased precision compared to integers and more straightforward implementation in Field Programmable Gate Array (FPGA) than the floating-point number system. In contemporary FPGA applications, rational fractions excel in compact hardware size, high throughput, and the ability to adjust precision through data width selection. The designed module for sigmoid activation function calculation demonstrates high throughput and occupies a relatively modest hardware volume compared to modules relying on piecewise polynomial approximation with fixed-point data.

ABSTRACT. Malware detection is an important topic of current cybersecurity, and Machine Learning appears to be one of the main considered solutions even if certain problems to generalize to new malware remain. In the aim of exploring the potential of quantum machine learning on this domain, our previous work showed that quantum neural networks do not perform well on image-based malware detection when using a few qubits. In order to enhance the performances of our quantum algorithms for malware detection using images, without increasing the resources needed in terms of qubits, we implement a new preprocessing of our dataset using Grayscale method, and we couple it with a model composed of five distributed quantum convolutional networks and a scoring function. We get an increase of around 20% of our results, both on the accuracy of the test and its F1-score.

ABSTRACT. This study explores the performance of Apple Silicon processors in real-world research tasks, with a specific focus on optimization and machine learning applications. Diverging from conventional benchmarks, various algorithms across fundamental datasets have been assessed using diverse hardware configurations, including Apple's M1 and M2 processors, NVIDIA RTX 3090 GPU and a mid-range laptop. The M2 demonstrates competitiveness in tasks such as breast cancer, liver and yeast classification, establishing it as a suitable platform for practical applications. Conversely, the dedicated GPU outperformed M1 and M2 on the eyestate1 dataset, underscoring its superiority in handling more complex tasks, albeit at the expense of substantial power consumption. With the technology advances, Apple Silicon emerges as a compelling choice for real-world applications, warranting further exploration and research in chip development. This study underscores the critical role of device specifications in evaluating machine learning algorithms.

ABSTRACT. The German Federal Criminal Police Office (BKA) reported damages of 72.6 million euros due to billing fraud in the German healthcare system in 2022, an increase of 25 % from the previous year. However, existing literature on automated healthcare fraud detection focuses on US, Taiwanese, or private data, and detection approaches based on individual claims are virtually nonexistent. In this work, we develop machine learning methods that detect fraud in German hospital billing data.

The lack of publicly available and labeled datasets limits the development of such methods. Therefore, we simulated inpatient treatments based on publicly available statistics on main and secondary diagnoses, operations and demographic information. We injected six different types of fraud that were identified from the literature. The simulation is validated using a private dataset from a big German health insurance and a human expert assessment. This is the first complete simulator for inpatient care data, enabling further research in inpatient care.

We trained and compared several Machine Learning models on the simulated dataset. Gradient Boosting and Random Forest achieved the best results with a weighted F1 measure of approximately 80 %. An in-depth analysis of the presented methods shows they excel at detecting compensation-related fraud, such as DRG upcoding. In contrast, other fraud types, such as early releases, are more difficult to detect from billing data alone. An impact analysis on private inpatient claims data of a big German health insurance company revealed that up to 12 % of all treatments were identified as potentially fraudulent.

ABSTRACT. This paper addresses a kernel-based learning problem for a network of agents locally observing a latent multidimensional, nonlinear phenomenon in a noisy environment. We propose a learning algorithm that requires only mild a priori knowledge about the phenomenon under investigation and delivers a model with corresponding non-asymptotic high probability error bounds. Both non-asymptotic analysis of the method and numerical simulation results are presented and discussed in the paper.

ABSTRACT. In the multi-disciplinary context of computational creativity and affective human-machine interaction, understanding and detecting creative processes accurately is advantage. This paper introduces a novel computational framework for creatively state detection, employing a multimodal approach that integrates emotions, arousal, and valence. The framework utilizes multimodal inputs to capture the creativity states, with emotion detection forming a foundational element. By fusioning emotions and emotional dimension, arousal, and valence. This paper outlines the theoretical foundations, key components, and integration principles of the proposed framework, paving the way for future advancements in computational creativity and affective computing.

ABSTRACT. Energy efficiency research aims to optimize the use of computing resources by minimizing energy consumption and increasing computational efficiency. This article explores the effect of the directive-based parallel programming model on energy efficiency for the multithreaded WZ factorization on multi-core central processing units (CPUs) and multi-core Graphics Processing Units (GPUs). Implementations of the multithreaded WZ factorization (both the basic and its variant optimized by strip-mining vectorization) are based on OpenMP and OpenACC. The evaluation of the proposed multithreaded implementations is made in terms of the energy efficiency of the WZ factorization for different sizes of input data on CPUs and GPUs. The implications of this research are manifold. The first is critical because strip-mining gave clear enhancement in comparison to the basic version. Second is, that the energy efficiency is much better on GPU than on CPU; and that on CPU, the use of OpenMP is more energy-efficient than the use of OpenACC; however, for GPU, OpenACC gives better results.

ABSTRACT. The purpose of the article is to develop a new dimensionality reduction algorithm for categorical data. We give a new geometric formulation of the PCA dimensionality reduction method for numerical data that can be effectively transferred to the case of categorical data with the Hamming metric. Numerical experiments confirm the effectiveness of the method.

ABSTRACT. An accurate and fast prediction of forest fire evolution is a crucial issue to minimize its impact. One of the challenges facing forest fire spread simulators is the uncertainty surrounding the input data. To mitigate this issue, directly measuring this data using sensors onboard Unmanned Aerial Vehicles (UAVs) in real time can significantly reduce its uncertainty. However, transmitting this data for processing and sending simulation results back to firefighters poses a challenge, specially in areas with low or no connectivity. To address this issue, we propose the use of Edge Computing. For simulation purposes, the FARSITE forest fires apread simulator has been used. This work aims to demonstrate the feasibility of leveraging Embedded Systems with low-consumption GPUs to simulate short-term forest fire spread evolution (up to 5 hours) at high resolution (5 meters). The capability of gathering data, simulating the hazard and delivering prediction results in situ, even without connectivity, could help decision makers to know about the wildfire potential just in few minutes.

ABSTRACT. We introduce our approach towards development of Digital Twins in Healthcare for both ischemic and haemorrhagic stroke, in relation to aetiology and prevention, treatment, and disease progression. These models start their development as generic Digital Twins and in a series of steps are strati-fied to become fully patient-specific. The example of a Digital Twin in Healthcare for treatment of ischemic stroke is described, and an outlook on the need and applicability of such digital twin in relation to stroke is provided.

ABSTRACT. Carbon nanotubes (CNTs) present immense promise for practical applications in chemical, electronics, and biomedical fields, owing to their unique mechanical, electrical, and structural properties [1]. However, a critical concern in utilizing CNTs in biomedical applications is their potential toxicity. Despite reports highlighting these adverse effects, a body of literature suggests either the absence or minimal toxicity associated with CNTs to healthy cells and tissues.[2] There are enough MD studies to the extent that confirm CNTs interaction with cellular membrane and some proteins.[3] however, there is no study investigating the interaction between dsDNA and swCNTs to study the mechanism of DNA breaking by swCNTs’ in genomics disorder or cancers. Small nanomaterials including swCNTs can be localized within the cell nucleus. CNTs may still come into contact with DNA when nuclear membrane breaks down during mitosis if they accumulate in cells but do not necessarily gain access to the nucleus. [4] This study utilizes molecular dynamics (MD) simulations to investigate the dynamic intricacies of the interaction between single-walled carbon nanotubes (swCNTs) and double-stranded DNA (dsDNA), complemented by docking results. By examining the influence of swCNT characteristics, as factors contributing to CNT toxicity such as length, radius, and chirality, our findings shed light on the complex interplay of binding affinity and stability of the DNA-CNT complex.[5] Comprehensive computational study including: spanning molecular docking results, binding energies, and MD simulation analyzing parameters provide a detailed understanding of dsDNA denaturation dynamics in expose of swCNT. The potential of mean force (PMF) profiles reveal the thermodynamic feasibility of the DNA-CNT interaction, outlining distinct energy landscapes and barriers. All the results confirmed the structural changes in dsDNA exposing to swCNT. Inter-molecular hydrogen bond reduction affirming dsDNA structural alterations as an unwinding two strains of DNA. Furthermore, the study indicates a significant alteration in the binding affinity of DNA to Ethidium Bromide as a representation of small molecules to dsDNA following it’s structural changes. This phenomenon could be generalized to another biomolecules - dsDNA interaction led to DNA’s dysfunctional disorders. This research offers valuable insights into the toxic effects of swCNTs on dsDNA, contributing to a rationalization of the cancerous potential of swCNTs.

References: [1] P.S. Karthik, A.L. Himaja, S.P. Singh, Carbon-allotropes: synthesis methods, applications and future perspectives, Carbon Lett. 15 (2014) 219e237. [2] K.A. Barcelos, J. Garg a , D.C. Ferreira Soares , A.L.Branco de Barros , Y. Zhao , La. Alisaraie : Review article, Recent advances in the applications of CNT-based nanomaterials in pharmaceutical nanotechnology and biomedical engineering, Journal of Drug Delivery Science and Technology 87 (2023) 104834 [3] M. Sahihi, Gh. Borhan, The effects of single-walled carbon nanotubes (SWCNTs) on the structure and function of human serum albumin (HSA): Molecular docking and molecular dynamic simulation studies, Springer Science+Business Media New York (2017) [4] P.D. Boyer, S.Ganesh, Z.Qin, B.D. Holt, M. J. Buehler, M.F. Islam, K.Noel Dahl : Delivering Single-Walled Carbon Nanotubes to the Nucleus Using Engineered Nuclear Protein Domains, ACS Applied Materials & Interfaces (2016) [5] D.Mohanta, S.Patnaik, S.Sood, N.Das: Carbon nanotubes: Evaluation of toxicity at biointerfaces: a review, Journal of Pharmaceutical Analysis 9 (2019) 293e300.

ABSTRACT. Augmented reality may soon revolutionize the world we live in. The incorporation of computer simulations into augmented reality glasses opens new perspectives for the perception of reality. In this paper, we investigate the possibility of performing numerical simulations in real time within augmented reality glasses. We present the technology that can be successfully employed in the real-life simulations of the Partial Differential Equations (PDE) based phenomena. We designed and implemented a two- and three-dimensional explicit dynamics solver in Lens Studio using Finite Difference Method (FDM) on the augmented reality glasses. We performed tests on the computational cost, memory usage, and the capability of performing real-life simulations of advection-diffusion and wave propagation problems.

ABSTRACT. Accurate prediction of the financial standing of companies remains a challenging task that primarily concerns stockholders, financial institutions, suppliers, investors, and researchers. Traditionally, the evaluation of a company’s financial standing is transformed into a binary classification, such as good or bad, or bankrupt or non-bankrupt. In fact, enterprises are also characterized by intermediate states, e.g., critical, poor, moderate, good, and excellent financial conditions. In this study, novel multi-class models are proposed to assess the financial condition of companies operating in nine countries in Central and Eastern Europe. Before starting the process of building a deep neural network model, the dataset is prepared and conditioned to eliminate any data discrepancy that could have a harmful effect on the model. Second, deep neural network models are trained for the evaluation of the financial standing of companies for defined classes. Third, the models are validated with a separate testing dataset. Finally, the stability of the models over time is evaluated for the 2015–2020 period (including the first year of the Coronavirus pandemic outbreak). A reduced binary class models were built to compare the accuracy of deep neural network models. The results demonstrate that the deep neural network models are stable over time and that the models for the Visegrad countries are characterised by the highest accuracy.

ABSTRACT. Thanks to the broad application fields, learning neural networks is still a more significant problem nowadays. Any attempt in the construction of faster learning algorithms is highly well come. This article presents a new way of learning neural networks with kernels with modified pseudo-inverse learning by modified SVD.

The new algorithm constructs the kernels during the learning and estimates the right number in the results. There is no longer a need to define the number of kernels before the learning. This means there is no need to test networks with a number of kernels that is too large, and the number of kernels is no longer a parameter in the selection process (in cross-validation).

The results show that the proposed algorithm constructs reasonable kernel bases, and final neural networks are accurate in classification.

ABSTRACT. We present experimental data and modelling results investigating the effects of plasma-assisted air cleaning systems on reducing the transmission of the SARS-CoV-2 virus among pupils in a primary school in Amsterdam, the Netherlands. We equipped 4 classrooms with 120 pupils with air cleaning systems, and 8 classrooms with approximately 240 pupils without such a system. We found a significantly lower number of infections in classrooms with air cleaning systems in the first two weeks after instalment, suggesting that air cleaning decreases aerosol transmission. In the subsequent weeks, however, infection numbers increased in the Netherlands, and the difference between classrooms with and without air cleaning ceased to be significant. We analyzed the experimental results, performed a Kaplan-Meier survival estimation and developed a SIR-based computational model that simulates the results of this experiment. We performed sensitivity analysis, optimised model parameters, and tested several hypotheses.

ABSTRACT. Breast cancer is a major public health problem, with 2.3 million new cases diagnosed worldwide each year. Improving immune system ability to fight cancer cells can prevent the development of several types of cancer and as such different immunotherapy techniques are being used according to several factors like histological grade and subtype of the tumours or the associated prognosis. However, the immune system depends on the local microenvironment to generate an effective response, and this requires regionally tailored trials, which results in a reduced number of patient samples. To minimise this drawback and improve the accuracy of patient prognosis predictions with machine learning models, we experiment in this work with several state-of-the-art data augmentation methods, i.e. noise injection, oversampling techniques and deep generative adversarial networks. The effectiveness of these methods is assessed through a set of immune system gene expression samples donated by 165 breast cancer patients from the Málaga region. Results showed a 5% increase in AUC and a 23 − 36% increase in F1 score for prognostic prediction associated with molecular subtype. Furthermore, the quality of the generated data and the subsequent improvement in prognostic prediction were supported by standard multivariate data analysis and distribution matching techniques.

ABSTRACT. This paper focuses on enhancing the lifespan of the Wireless Sensor Networks by integrating a distributed Learning Automaton into its operation. The proposed framework seeks to determine an optimized activity schedule that extends the network's lifespan while ensuring that the monitoring of designated target areas meets predefined coverage requirements. The proposed algorithm harnesses the advantages of localized algorithms, including leveraging limited knowledge of neighboring nodes, fostering self-organization, and effectively prolonging the network's longevity while maintaining the required coverage ratio in the target field.

ABSTRACT. The paper proposes coupling Parametric Integral Equation System (PIES) and Physics-Informed Neural Network (PINN) for solving two-dimensional boundary value problems. As a result, the computational domain can be decomposed into subdomains, where solutions are obtained independently using PIES and PINN while simultaneously satisfying interface connection conditions. The efficacy of this approach is validated through a numerical example illustrating its accuracy in solving potential problems defined by Laplace's equation.

ABSTRACT. Product units are components of neural networks that can learn higher-order terms and have a higher empirical information capacity than classical summation units. For instance, it has been shown empirically that all logical functions of two binary parameters can be represented by single product units, which is impossible with a linear perceptron that fails to represent e.g. the XOR function. The XOR problem can be generalized to higher dimensions as the parity problem that is also known to be solvable by a single product unit. However, learning the parity function on product units with local optimization algorithms such as stochastic gradient descent often fails due to the occurrence of local minima of the loss function, inspiring work on global optimization methods for product unit networks. Further, to the best of our knowledge, the trainability and limitations of product units have not been investigated with analytical methods yet. In this paper we explore several modifications of product units that both improve the local convergence properties and allow analytical derivations of some of the results reported in the literature. We test the methods also with noisy input data and make use of the extrapolation capabilities of product units using incomplete data sets for training. Finally, we explore the representability of functions beyond the parity function with geometric methods. In particular, we show that the n-dimensional disjunction OR can be learned with a single product unit, while this is impossible for the conjunction AND.

ABSTRACT. Inspired by Lyapunov control techniques for quantum systems, feedback-based quantum algorithms have recently been proposed as alternatives to variational quantum algorithms for solving quadratic unconstrained binary optimization problems. These algorithms update the circuit parameters layer-wise through feedback from measuring the qubits in the previous layer to estimate expectations of certain observables. Therefore, the number of samples directly affects the algorithm's performance and may even cause divergence. In this work, we propose an adaptive technique to mitigate the sampling noise by adopting a switching control law in the design of the feedback-based algorithm. The proposed technique can lead to better performance and convergence properties. We show the robustness of our technique against sampling noise through an application for the maximum clique problem.

ABSTRACT. In recent years, immune checkpoint inhibitors (ICIs)–developed primarily from the knowledge of T-cell immunoreceptor (TCR) signaling have revolutionized the treatment of various cancers, offering new-found hope to patients. However, their usage is associated with a considerable incidence of immune-related adverse events (irAEs). While ICI-induced blockage of TCR inhibitory signaling pathways can successfully bolster anti-tumor immune response, it remains that as much as 40% of patients suffer irAEs in some capacity during treatment. These adverse effects range from manageable conditions like gastrointestinal irregularity to severe, life-threatening complications such as colitis and pneumonitis. With the delay of immunotherapy occurring in up to 73% of patients as a result of irAEs, accurate phenotyping, and early detection of these adverse effects are critical for timely intervention and optimal patient management. To rise to this pursuit, we introduce the Clinical Analysis and Reporting Enhancement Network, or CARENET. A HIPAA-compliant proactive monitoring aid in clinical environments, CARENET utilizes advancements in Natural Language Processing (NLP) and knowledge representation to provide electronic health record (EHR) data-driven insights for commencing or in-progress ICI treatments. By utilizing vast repositories of data and providing answers to practitioner inquiries via powerful, detailed human-computer interaction (HCI), CARENET improves patient outcomes by enabling faster, more informed clinical treatment decisions.

ABSTRACT. Survival analysis focuses on the prediction of failure time and serves as an important prognostic tool, not solely confined to medicine but also across diverse fields. Machine learning methods, especially decision trees, are increasingly replacing traditional statistical methods which are based on assumptions that are often difficult to meet. The paper presents a new global method for inducing survival trees containing Kaplan--Mayer estimators in leaves. Using a specialized evolutionary algorithm, the method searches for oblique trees in which multivariate tests in internal nodes divide the feature space using hyperplanes. Specific variants of mutation and crossover operators have been developed, making evolution effective and efficient. The fitness function is based on the integrated Brier score and prevents overfitting taking into account the size of the tree. A preliminary experimental verification and comparison with classical univariate trees was carried out on real medical datasets. The evaluation results are promising.

ABSTRACT. The recent history of society has seen changes in the way and places where peo-ple live for leisure, work and/or study, due to the influence of various factors, in-cluding economic, geographic, and even public health. The fact that the popula-tion increasingly lives in closed construction structures or in increasingly urban-ized locations, where concrete is the predominant element, has sparked research-ers' interest in knowing how these changes can influence their behavior and health. Several studies indicate that humans have an innate connection with nature and that these environments have a restorative effect on mental and attention fa-tigue, helping to prevent stress and revitalize well-being. Given the relevance of the above, this investigation intends to delve deeper into this topic based on the proposal to apply an existing methodology adapted for experimenting with living spaces plus natural elements simulated in virtual reality (VR), simultaneously monitored by physiological, behavioral, and psychological aspects. of the partici-pants. It is expected that the implementation of the methodology, by providing an immersive experience accompanied by real-time monitoring and a multi-method analysis approach from three perspectives, will allow the identification of con-sistent contributions from the incorporation of biophilia in environments that are part of people's routine, obtaining subsidies important so that professionals such as architects, simulation developers, psychologists, public and business managers can invest adequate efforts and resources that effectively reflect a better living condition at work or leisure for the population.

ABSTRACT. The understanding of large-scale scientific software poses significant challenges due to its diverse codebase, extensive code length, and target computing architectures. The emergence of generative AI, specifically large language models (LLMs), provides novel pathways for understanding such complex scientific codes. This paper presents S3LLM, an LLM-based framework designed to enable the examination of source code, code metadata, and summarized information in conjunction with textual technical reports in an interactive, conversational manner through a user-friendly interface. In particular, S3LLM utilizes open-source LLaMA-2 models to improve code analysis by converting natural language queries into Feature Query Language (FQL) queries, facilitating the quick scanning and parsing of entire code repositories. In addition, S3LLM is equipped to handle diverse metadata types, including DOT, SQL, and customized formats. Furthermore, S3LLM incorporates retrieval augmented generation (RAG) and LangChain technologies to directly query extensive documents. S3LLM demonstrates the potential of using locally deployed open-source LLMs for the rapid understanding of large-scale scientific computing software, eliminating the need for extensive coding expertise, and thereby making the process more efficient and effective. S3LLM is available at https://github.com/ResponsibleAILab/s3llm

ABSTRACT. A rigorous physics-informed learning methodology is proposed for predictions of wave solutions and band structures in electronic and optical superlattice structures. The methodology is enabled by proper orthogonal decomposition (POD) and Galerkin projection of the wave equation. The approach solves the wave eigenvalue problem in POD space constituted by a finite set of basis functions (or POD modes). The POD ensures that the generated modes are optimized and tailored to the parametric variations of the system. Galerkin projection however enforces physical principles in the methodology to further enhance the accuracy and efficiency of the developed model. It has been demonstrated that the POD-Galerkin methodology offers an approach with a reduction in degrees of freedom by 4 orders of magnitude, compared to direct numerical simulation (DNS). A computing speedup near 15,000 times over DNS can be achieved with high accuracy for either of the superlattice structures if only the band structure is calculated without the wave solution. If both wave function solution and band structure are needed, a 2-order reduction in computational time can be achieved with a relative least square error (LSE) near 1%. When the training is incomplete or the de-sired eigenstates are slightly beyond the training bounds, an accurate prediction with an LSE near 1%-2% still can be reached if more POD modes are included. This reveals its remarkable learning ability to reach correct solutions with the guidance of physical principles provided by Galerkin projection.

ABSTRACT. In infectious disease models, the mechanism used to propagate infection in the simulated population is crucial to the accuracy of the model. In this paper, we compare some of the popular disease models used to model the COVID-19 pandemic in terms of the mechanism they use to propagate infection in the simulated population. Our analysis shows that the basic reproduction number $R_0$ plays a crucial role in the transmission of the disease in the studied models. Based on certain limitations of the $R_0$ parameter, we propose an improvement to the existing disease transmission formulations. We argue that the proposed formulation is more generalizable to different regions and different stages of the disease and makes the model adjust to the changing population structure and behavior.

ABSTRACT. At RIKEN, we have been exploring a fusion of big data and big computation in numerical weather prediction (NWP), and now with AI and machine learning (ML). Our group in RIKEN has been pushing the limits of NWP through two orders of magnitude bigger computations using the previous Japan’s flagship “K computer”. The efforts include 100-m mesh, 30-second update “Big Data Assimilation” (BDA) fully exploiting the big data from a novel Phased Array Weather Radar. With the new “Fugaku” since 2021, we achieved a real-time BDA application to predict sudden downpours up to 30 minutes in advance during Tokyo Olympics and Paralympics. This presentation will introduce the most recent results from BDA experiments, followed by perspectives toward DA-AI fusion and expanding new applications beyond meteorology..

ABSTRACT. Virtual planning tools that provide intuitive user interaction and immediate hemodynamic feedback are crucial for cardiologists to effectively treat coronary artery disease. Current FDA-approved tools for coronary intervention planning require days of preliminary processing and rely on conventional 2D displays for hemodynamic evaluation. Immersion offered by extended reality (XR) has been found to benefit intervention planning over traditional 2D displays. To bridge these gaps, we introduce HarVI, a coronary intervention planner that leverages machine learning for real-time hemodynamic analysis and extended reality for intuitive 3D user interaction. The framework uses a predefined set of 1D computational fluid dynamics (CFD) simulations to perform one-shot training for our machine learning-based blood flow model. In a cohort of 50 patients, we calculated fractional flow reserve (FFR), the gold standard biomarker of ischemia in coronary disease, using HarVI (FFR-HarVI) and 1D CFD models (FFR-1D). HarVI was shown to almost perfectly recapitulate the results of 1D CFD simulations through continuous and categorical validation scores. In this study, we establish a machine learning-based process for virtual coronary treatment planning with an average turnaround time of just 74 minutes, thus reducing the required time for one-shot training to less than a working day.

ABSTRACT. Monitoring disease development in the coronary arteries, which supply blood to the heart, is crucial and can be assessed via hemodynamic metrics. While these metrics are known to depend on the inlet velocity, the effects of changes in the time-dependent inlet flow profile are not understood. In this study, we seek to quantify the effects of modulating temporal arterial waveforms to understand the effects of hemodynamic metrics. We expand on previous work that identified the minimum number of points of interest needed to characterize a left coronary artery inlet waveform. We vary these points of interest and quantify the effects on commonly used hemodynamic metrics such as wall shear stress, oscillatory shear index, and relative residence time. To simulate we use 1D Navier-Stokes and 3D lattice Boltzmann simulation approaches conducted on high performance compute clusters. The results allow us to observe which parts of the waveform are most susceptible to perturbations and, therefore also to measurement error. The impacts of this work include clinical insight as to which portions of velocity waveforms are most susceptible to measurement error, the construction of a method that can be applied to other fluid simulations with pulsatile inlet conditions, and the ability to distinguish the vital parts of a pulsatile inlet condition for computational fluid dynamic simulations.

ABSTRACT. Chronic Obstructive Pulmonary Disease (COPD) is an advancing respiratory condition marked by sustained airflow limitation in the pulmonary and stands as the third leading cause of morbidity and mortality on a global scale. While the healthcare system trying to control and manage the severity of the disease, a significant number of patients are still directed daily to the Emergency Department (ED), impacting both patients and the healthcare system. In this project we implement a simulation of Real Clinical Trials, in other words, Virtual Clinical Trials (VCT), which can provide a series of possibilities that real clinical trials could not offer, such as patients with different percentages of adherence to treatment, creation of virtual cohorts with specific characteristics and selection of environment conditions of the trial. We illustrate a cohort-based management strategy that utilizes large-scale data analytics to identify patterns and trends within the COPD patient population. Also, we advocate the adoption of a finite-state machine approach for modeling COPD exacerbation, intending to refine our comprehension of COPD patient conditions in the context of cohort monitoring. Further research and validation are essential to refine and scale this integrated model

ABSTRACT. Billions of people work every week. Forecasting which tasks gets done in a working week is important because it could help workers understand (i) whether their workload is manageable and (ii) which task scheduling approach helps to maximize the amount of work done and/or minimize the negative consequences of unfinished work. Here we present a working week simulation prototype, R2, and showcase how it can be used to forecast the working week for three archetypical workers. We show that R2 forecasts are sensitive to different task loads, task scheduling strategies and different levels of emerging work complications. We also highlight how R2 supports a new type of validation setting, namely that of user self-validation, and discuss the advantages and drawbacks of this new validation approach. We provide R2 as an online platform to allow users to create their own worker profile and task lists, and believe the tool could serve as a starting point for more in-depth research efforts on user-centric working week modelling.

ABSTRACT. Efficient route pruning algorithms play a critical role in movement simulation scenarios, particularly in migration models within regions with limited infrastructure, such as South Sudan. This paper presents a novel approach that incorporates health facility density into route pruning algorithms, with the aim of enhancing the accuracy of route networks generated by such algorithms. The underlying assumption is that the distribution of healthcare facilities is pivotal for refugees, and accessibility to such services influences their route selection towards safe locations. Drawing from the challenges faced in South Sudan’s routes, characterized by vast distances, poor road conditions, and sparse health infrastructure, we present an extension to existing route pruning algorithms tailored to address these specific challenges. The algorithm accounts for the distribution and density of health facilities along potential routes, ensuring that selected paths prioritize areas with higher healthcare needs while maximizing efficiency. Our methodology involves leveraging geographic information systems data to map the locations of existing health facilities and population centres. We then adapt the route pruning algorithm to consider not only distance but also the availability and proximity of health facilities along candidate routes. To validate the effectiveness of our approach, we conduct a case study in South Sudan and compare the results with the previous algorithm. We expect that the findings will demonstrate the superior performance of this algorithm in improving the average relative difference over the previous algorithm. This aims to improve healthcare logistics and accessibility in similar scenarios.

ABSTRACT. On the path toward developing the first fusion energy devices, plasma simulations have become indispensable tools for supporting the design and development of fusion machines. Among these critical simulation tools, BIT1 is an advanced Particle-in-Cell code with Monte Carlo collisions, specifically designed for modeling plasma-material interaction and, in particular, analyzing the power load distribution on tokamak divertors. The current implementation of BIT1 relies exclusively on MPI for parallel communication and lacks support for GPUs. In this work, we address these limitations by designing and implementing a hybrid, shared-memory version of BIT1 capable of utilizing GPUs. For shared-memory parallelization, we rely on OpenMP and OpenACC, using a task-based approach to mitigate load-imbalance issues in the particle mover. On an HPE Cray EX computing node, we observe an initial performance improvement of approximately 42%, with scalable performance showing an enhancement of about 38% when using 8 MPI ranks. Still relying on OpenMP and OpenACC, we introduce the first version of BIT1 capable of using GPUs. We investigate two different data movement strategies: unified memory and explicit data movement. Overall, we report BIT1 data transfer findings during each PIC cycle. Among BIT1 GPU implementations, we demonstrate performance improvement through concurrent GPU utilization, especially when MPI ranks are assigned to dedicated GPUs. Finally, we analyze the performance of the first BIT1 GPU porting with the NVIDIA Nsight tools to further our understanding of BIT1’s computational efficiency for large-scale plasma simulations, capable of exploiting current supercomputer infrastructures.

ABSTRACT. Computational electromagnetics plays a crucial role across diverse domains, notably in fields scuh as antenna design and radar signature prediction, owing to the omnipresence of electromagnetic phenomena. Numerical methods have replaced traditional experimental approaches, expediting design iterations and scenario characterization. The emergence of GPU accelerators offers an efficient implementation of numerical methods that can significantly enhance the computational capabilities of partial differential equations (PDE) solvers with specific boundary-value conditions. This paper explores parallelization strategies for implementation of a Finite-Difference Time-Domain (FDTD) solver on GPUs, leveraging shared memory and optimizing memory access patterns to achieve performance gains. One notable innovation presented in this research involves the utilization of strategies such as exploiting temporal locality and avoiding misaligned global memory ccesses to enhance data processing efficiency. Additionally, we break down the computation process into multiple kernels, each focusing on computing different components of the electromagnetic (EM) field, to enhance shared memory utilization and GPU cache efficiency. We implement crucial design optimizations to fully exploit GPU's parallel processing capabilities, including maintaining consistent block sizes, analyzing optimal configurations for field-updating kernels, and optimizing memory access patterns for CUDA threads within warps. Our experimental analysis verifies the effectiveness of these strategies, resulting in improvements in both reducing execution time and enhancing the GPU's effective memory bandwidth. Throughput evaluation demonstrates performance gains, with our CUDA implementation achieving up to 17 times higher throughput compared to CPU-based methods. Speedup gains and throughput comparisons illustrate the scalability and efficiency of our approach, showcasing its potential for developing large-scale electromagnetic simulations on GPUs.

ABSTRACT. As data sizes continue to expand, the challenge of conducting meaningful analysis grows. Utilizing I/O (Input/Output) libraries, such as HDF5 (Hierarchical Data Format) and ADIOS2 (Adaptable IO System), facilitates the filtering of raw data, with prior research highlighting the advantages of dissecting these formats for enhanced metadata management. Our study introduces a novel data management technique aimed at boosting query performance for HPC analysis applications through the automatic precomputation of commonly used data characteristics, as identified by our user survey. The Data Analysis Interface (DAI), developed for the JULEA storage framework, not only enables querying this enriched metadata but also shows how domain-specific features can be integrated, demonstrating a potential improvement in query times by up to 22,000 times.

ABSTRACT. Generally, evaluating the performance offered by an HPC I/O system with different configurations and the same application allows selecting the best settings. Both users and researchers often want to make changes to this type of system to analyze how this affects their applications. By testing configurations in a simulated environment, the risk of causing disruptions or damage to real systems is reduced. This paper proposes to use agent-based modeling and simulation (ABMS) to evaluate the performance of the I/O software stack. In particular, the modeling of the communications layer of the PVFS2 file system is presented in detail. ABMS has been selected because it enables a rapid change in the level of analysis for modeling and implementing a simulator. It can focus on both macro- and micro-levels (how the aggregate behavior of system agents is born and analyzing their individual behavior).

ABSTRACT. The design and implementation of algorithms for increasingly large and complex modern supercomputers requires the definition of data structures and workload distribution mechanisms in a productive and scalable way. In this paper, the focus is on parallel tree-based algorithms that explore unbalanced trees using the Depth-First Search (DFS) strategy. Due to their irregular and unpredictable nature, such algorithms raise multiple challenges, such as dynamic load balancing. We propose a PGAS data structure and work-stealing mechanism for this class of algorithms. It consists in a multi-pool accessed in the depth-first order and the work-stealing takes place when a given worker gets out of work. Their PGAS-based design is driven by the higher level of abstraction and scalability. Their implementation is based on the Chapel language and is provided in the open-source module called DistributedBag_DFS. For experimental evaluation, four distributed applications are provided: two backtracking to solve instances of the Unbalanced Tree-Search benchmark and N-Queens problem, and two branch-and-bound implementations to solve to the optimality instances of the permutation Flowshop Scheduling and 0/1-Knapsack problems. According to the experimental results, it is shown that the data structure and work-stealing scheme allow the applications to scale in both shared- and distributed-memory settings, achieving 97% and 94% of the ideal speed-up for its best results, respectively. Furthermore, in the larger scale experiments, where up to 400 computer nodes (51,200 processing cores) are used, 50% of strong scaling efficiency is achieved, showcasing the robustness and efficiency of our approach.

ABSTRACT. A straightforward way of solving global optimization problems is to find all local optima of the objective function. Therefore, the ability of detecting multiple local optima is a key feature of a practically usable global optimization method. One of such methods is a multi-population evolutionary strategy called the Hierarchic Memetic Strategy (HMS). Although HMS has already proven its global optimization capabilities there is an area for improvement. In this paper we show such an enhancement resulting from the application of the Nearest-Better Clustering. Results of experiments consisting both of curated benchmarks and a real-world inverse problem show that on average the performance is indeed improved compared to the baseline HMS and remains on par with state-of-the-art evolutionary global optimization methods.

ABSTRACT. Physics-informed neural networks (PINNs) provide a means of obtaining approximate solutions of partial differential equations and systems through the minimisation of an objective function which includes the evaluation of a residual function at a set of collocation points within the domain. The quality of a PINNs solution depends upon numerous parameters, including the number and distribution of these collocation points. In this paper we consider a number of strategies for selecting these points and investigate their impact on the overall accuracy of the method. In particular, we suggest that no single approach is likely to be “optimal” but we show how a number of important metrics can have an impact in improving the quality of the results obtained when using a fixed number of residual evaluations. We illustrate these approaches through the use of two benchmark test problems: Burgers’ equation and the Allen-Cahn equation.

ABSTRACT. In this paper we consider a training of Physics Informed Neural Networks with fully connected neural networks for approximation of solution of one-dimensional advection-diffusion problem. In this context, the neural network is interpreted as a non-linear functions of one spatial variable, approximating the solution scalar field, namely $y=PINN(x)=A_n \sigma(A_{n-1} ...A_2\sigma(A_1+b1)+b2)+...+b_{n-1})+b_n$. In standard PINN approach, the $A_i$ denotes dense matrices, $b_i$ denotes bias vectors, and $\sigma$ is the non-linear activation function (sigmoid in our case). In our paper, we consider a case when $A_i$ are hierarchical matrices $A_i={\cal H}_i$. We assume a structure of our hierarchical matrices approximating the structure of finite difference matrices employed to solve analogous PDEs. In this sense, we propose a hierarchical neural network for training and approximation of PDEs using PINN method. We verify our method on the example of one-dimensional advection-diffusion problem.

ABSTRACT. We propose and assess a novel deep learning approach for solving partial differential equations (PDEs) with discontinuities in the data and/or the solution. To better capture the discontinuities, we draw inspiration from the classical discontinuous Galerkin method. Emulating the finite element method, we partition the domain and approximate the PDE solution within each element using polynomials of a specified degree. We construct a neural network to compute the solution’s degrees of freedom. We propose a loss function based on the quadratic Ritz functional, that we correct with jump terms to enforce the solution and flux continuity on the mesh skeleton. In the optimization process, we employ exact (Gaussian) quadrature rules and automatic differentiation.

In this talk, we will present the methodology, its implementation, and performance. We will assess its efficiency by experimenting with different architectures, optimizers, and weights in the loss function with a simple 1D problem. Motivated by the ADAM optimizer’s sensitivity to hyperparameters, convergence issues, and limited theoretical justifications, we will consider a hybrid optimization strategy, in which the optimal parameters of the last layer are initially determined using a least-squares solver in a minimum-residual sense. To weight the different terms in the loss functions, we draw inspiration from the standard values employed in the discontinuous Galerkin method. We will also explore alternative weighting schemes suggested in the literature. We will assess the method’s efficiency against the two options. Then, we will show how the proposed method performs with several two-dimensional boundary value problems with discontinuities in the data and/or the solution.

ABSTRACT. Over the last two decades, Goal-Oriented Adaptivity (GOA) has been extensively studied and developed for the Finite Element Method (FEM). It is a technique that consists of enhancing the discretization of the underlying mesh to approximate a specific Quantity of Interest (QoI) rather than the error in the energy norm.

In this work, instead of considering a FEM discretization, we use neural networks. By adopting an Extreme Learning Machine interpretation for feed-forward neural networks, we are able to establish a Petrov-Galerkin-type approach such that: (i) the basis functions are no longer locally supported, and thus possess the potential to overcome the curse of dimensionality; (ii) the basis functions are governed by trainable parameters, yielding a highly non-linear scheme; and (iii) for each choice of trainable parameters, the optimal coefficients of the corresponding linear combination are efficiently computable following a minimum residual methodology. In this way, each GOA iteration is identified in our proposal as a training step of neural networks where the loss function represents an appropriate upper bound of the QoI.

We restrict ourselves to symmetric and positive-definite problems to ensure that we can incorporate robust error estimators for the primal and dual problems within the loss function. Extending it to non-symmetric or indefinite problems requires special care, which we plan to address in future work. The numerical experiments conducted in different spatial dimensions demonstrate the effectiveness of our strategy, up to the optimizer performance capacity and the usual numerical integration challenges.

ABSTRACT. This paper deals with the restricted longest common subsequence (RLCS) problem, an extension of the well-studied longest common subsequence problem involving two sets of strings: the input strings and the restricted strings. This problem has applications in bioinformatics, particularly in identifying similarities and discovering mutual patterns and motifs among DNA, RNA, and protein molecules. We introduce a general search framework to tackle the RLCS problem. Based on this, we present an exact best-first search algorithm and a metaheuristic Beam Search algorithm. To evaluate the effectiveness of these algorithms, we compare them with two exact algorithms and two approximate algorithms from the literature along with a greedy approach. Our experimental results show the superior performance of our proposed approaches. In particular, our exact approach outperforms the other exact methods in terms of significantly shorter computation times, often reaching an order of magnitude compared to the second-best approach. Moreover, it successfully solves all problem instances, which was not the case with the other approaches. In addition, Beam Search provides closeto- optimal solutions with remarkably short computation times.

ABSTRACT. Electronic design automation toolchains require solving various circuit manipulation problems, such as floor planning, placement and routing. These circuits may be implemented using either Very Large-Scale Integration (VLSI) or Field Programmable Gate Arrays (FPGAs). However, with the ever-increasing size of circuits, now up to billions of gates, straightforward approaches to these problems do not scale well. A possible approach to reduce circuit complexity is to cluster circuits, to reduce their apparent size for critical processing operations, while preserving their topological properties (e.g., connection locality).

Several models have been proposed to tackle the clustering problem. In this work, we consider the problem of clustering combinatorial circuits, without cell replication. Our main objective is to minimize the overall delay, which conditions the circuit operating frequency. We propose a dedicated clustering algorithm based on binary search and study and improve the existing parameterized approximation ratio from M^2+M (with M being the maximum size of each cluster) to M under specific hypothesis. We present an extension of the weighting schemes introduced to model path length more accurately. This weighting scheme is combined with clustering methods based on a recursive matching algorithm. We evaluate and compare our approximation algorithm and recursive matching on several circuit instances and we obtain better results for a large number of instances with our algorithm than recursive matching.

ABSTRACT. Optimal mesh refinement is important for finite element simulations, facilitating the generation of non-uniform meshes. While existing neural network-based approaches have successfully generated high quality meshes, they can only handle a fixed number of vertices seen during training. We introduce GraphMesh, a novel mesh refinement method designed for geometric generalization across meshes with varying vertex counts. Our method employs a two-step process, initially learning a unified embedding for each node within an input coarse mesh, and subsequently propagating this embedding based on mesh connectivity to predict error distributions. By learning a node-wise embedding, our method achieves superior simulation accuracy with reduced computational costs compared to current state-of-the-art methods. Through experimentation and comparisons, we showcase the effectiveness of our approach across various scenarios, including geometries with different vertex counts. We validated our approach by predicting the local error estimates for the solution of Poisson’s equation.

ABSTRACT. The aim of this work was to design an algorithm for rendering volumetric clouds in real time using a voxel representation. The results were verified using reference renders created with the Blender program using the Principled Volume shader. The important properties of the algorithm that were tried to be achieved are the ability to display clouds with different characteristics (thin and dense clouds) and the speed of operation enabling interactivity. We proposed a method consisting of two parameterizable image display algorithms with various performance and properties. The starting point was the single-scattering algorithm, which was extended with precalculation, and a simplified form of multi-scattering. Individual methods will be compared with reference images. Methods performing similar tasks, depending on the purpose, generate single image frames at a rate ranging from several dozen hours to a few seconds. Using the described mechanisms, the proposed method allowed to achieve times between 1 and 200 milliseconds, depending on the method variant and quality settings.

ABSTRACT. Recently proposed gradient estimators enable gradient descent over stochastic programs with discrete jumps in the response surface, which are not covered by automatic differentiation (AD) alone. Although these estimators' capability to guide a swift local search has been shown for certain problems, their applicability to models relevant to real-world applications remains largely unexplored. As the gradients governing the choice in candidate solutions are calculated from sampled simulation trajectories, the optimization procedure bears similarities to metaheuristics such as particle swarm optimization, which puts the focus on the different methods' calibration progress per function evaluation. Here, we consider the calibration of force-based crowd evacuation models based on the popular Social Force model augmented by discrete decision making. After studying the ability of an AD-based estimator for branching programs to capture the simulation's rugged response surface, calibration problems are tackled using gradient descent and two metaheuristics. As our main insights, we find 1) that the estimation's fidelity benefits from disregarding jumps of large magnitude inherent to the Social Force model, and 2) that the common problem of calibration by adjusting a simulation input distribution obviates the need for AD across the Social Force calculations, allowing gradient descent to excel.

ABSTRACT. Using a minimal but sufficient closed-loop encapsulation and the theoretical framework of classical control, we implement and test the mathematical model of the baroregulation due to Mauro Ursino [21]. We present and compare data from a local relative sensitivity analysis and an input parameter orthogonality analysis from an unregulated and then an equivalent, regulated cardiovascular model with a single ventricle and “CRC” Windkessel representation of the systemic circulation. We conclude: (i) a basic model of the closed loop control is intrinsically stable; (ii) regulation generally (but not completely) suppresses the sensitivity of output responses on mechanical input parameters; (iii) with the sole exception of the regulation set-point, the mechanical input parameters are more influential on system outputs than the regulation input parameters.

ABSTRACT. Abstract. This research investigates how collectives of doctors influence their diagnostic method preferences within small-world network social struc-tures through participation in diverse types of medical practice-sharing activi-ties across different scales. We propose an approach based on vectorization of the preferences for various diagnostic methods among physicians, quanti-fying their openness to these methods using the Shannon diversity index. Utilizing theoretical foundations from threshold models, influence models, and the Hegselmann-Krause model, we designed simulation experiments for teaching activities and seminars to explore the dynamic changes in prefer-ence vectors and Shannon diversity indices among these doctors in a small-world network. We evaluated our approach with a real-world data set on ver-tigo treatment by several clinical specialists of different specialty (neurolo-gists, otolaryngologist). Building on real data from this initial group, we then simulated data for a large number of doctors from various medical commu-nities to examine phenomena in larger-scale systems. Hierarchical networks featuring small-world properties were developed to simulate “local” within-community and “global” across-community seminars, reflecting different in-tra- and inter-community scenarios. The experiments show different patterns of practice converging during simulation in various scales and scenarios. The findings of this study provide significant insights for further research into practice-based knowledge sharing among healthcare professionals, highlight-ing the nuanced interplay between social network structures and professional consensus formation.

ABSTRACT. The aim of the research is to assess the applicability of methods of artificial intelligence to the analysis and prediction of infectious disease dynamics, with an aim to increase the speed of obtaining predictions along with enhancing quality of the results. To ensure the compliance of the forecasts with the natural laws governing the epidemic transmission, we employ Physics-Informed Neural Networks (PINN) as our main tool for the forecasting experiments. PINN is used for estimating coefficients of the SIRD epidemiological model implemented in the form of the system of ordinary differential equations. We compare the accuracy of different implementations of PINN along with the statistical models in the task of forecasting COVID incidence in Saint Petersburg, thus choosing the best modeling approach for this challenge. With the help of numerical experiments, we show the applicability of the approach to infectious disease modeling based on coupling classic approaches, namely, SIR models, and the cutting-edge research related to machine learning techniques. The results of the research could be incorporated into surveillance systems monitoring the advance of COVID and influenza incidence in Russian cities.

ABSTRACT. The application of computational science methods and tools in healthcare is growing rapidly. These methods support decision-making and policy development. They are commonly used in decision support systems (DSSs) used in many fields. This paper presents a decision support system based on the newly developed SSP-SPOTIS (Strong Sustainable Paradigm based Stable Preference Ordering Towards Ideal Solution) method. The application of the proposed DSS is demonstrated in the example of assessing healthcare systems of selected countries concerning resilience to pandemic-type crisis phenomena. The developed method considers the strong sustainability paradigm by reducing linear compensation criteria with the possibility of its modeling. The research demonstrated the usefulness, reliability, and broad analytical opportunities of DSS based on SSP-SPOTIS in evaluation procedures focused on sustainability aspects considering a strong sustainability paradigm.

ABSTRACT. Medical data often presents as a time series, reflecting the disease's progression. This can be captured through longitudinal health records or hospital treatment notes, encompassing diagnoses, health states, medications, and procedures. Understanding disease evolution is critical for effective treatment. Graph embedding of such data is advantageous, as it inherently captures entity relationships, offering significant utility in medicine. Hence, this study aims to develop a graph representation of Electronic Health Records (EHRs) and combine it with a method for predictive analysis of COVID-19 using network-based embedding. Evaluation of Graph Neural Networks (GNNs) against Recurrent Neural Networks (RNNs) reveals superior performance of GNNs, underscoring their potential in medical data analysis and forecasting.

ABSTRACT. Deep neural network models are commonly used in computer vision problems, i.e. for an image segmentation task. Convolutional neural networks have been state-of-the-art methods in image processing, but new architectures, such as Transformer-based approaches, have started outperforming previous techniques in many applications. However, those techniques are still not commonly used in urban analyses, mostly performed manually. This paper presents a framework for the residential building semantic segmentation architecture as a tool for automatic urban phenomena monitoring. The method could improve urban decision-making processes with automatic city analysis, which is predisposed to be faster and even more accurate than those made by human researchers. The study compares the application of new deep network architectures with state-of-the-art solutions. The analysed problem is urban functional zone segmentation for the urban sprawl evaluation using targeted land cover map construction. The proposed method monitors and rates the expansion of the city, which, uncontrolled, can cause adverse effects. The method was tested on orthophotos from three residential districts. The first district has been manually segmented by functional zones and used for model training and evaluation. The other two districts have been used for automated segmentation by models' inference to test the robustness of the methodology. The test resulted in 98.2% efficiency.

ABSTRACT. Travel mode choice models make it possible to learn under what conditions people decide to use different means of transport. Typically, such models are based on real trip records provided by respondents, e.g. city inhabitants. However, the question arises of how to scale the insights from an inevitably limited number of trips described in their travel diaries to entire cities. To address the limited availability of real trip records, we propose the Urban Journey System integrating big data platforms, analytic engines, and synthetic data generators for urban transport analysis. First of all, the system makes it possible to generate random synthetic journeys linking origin and destination pairs by producing location pairs using an input probability distribution. For each synthetic journey, the system calculates candidate routes for different travel modes (car, public transport (PT), cycling, and walking). Next, the system calculates Level of Service (LOS) attributes such as travel duration, waiting time, and walking distances involved, assuming both planned and real behaviour of the transport system. This allows us to compare travel parameters for planned and real transits. We validate the system with spatial, schedule and GPS data from the City of Warsaw. We analyse LOS attributes and underlying vehicle trajectories over time to estimate spatio-temporal distributions of features such as travel duration, and number of transfers. We extend this analysis by referring to the travel mode choice model developed for the city.

ABSTRACT. The application of Internet of Things technology in the agricultural sector has allowed to achieve a significant improvement in the process of growing and harvesting products. This has been possible since it allows obtaining a more exact control of the information in real time, thus allowing better deci-sion-making in crop management and thereby improving their quality. Faced with this situation, this paper proposes the design and implementation of a soil moisture monitoring system for a strawberry crop using LoRaWAN technology to allow the farmer to improve the production of their crops, while maintaining low technological implementation costs. The system al-lows the visualization of the data in real time, which are obtained from the sensors installed in the ground and which are transmitted through the Lo-RaWAN network. Once the system was developed using different trending technological tools, its functionality could be verified with satisfactory re-sults. The functionality of the application obtained an acceptance of 94% and an usability a score of 86.87, indicating that the system meets the expec-tations of the users. Additionally, in the coverage tests, it was possible to verify the long communication range of the installed LoRaWAN devices.

ABSTRACT. In the domain of battery energy storage systems for Electric Vehicles (EVs) applications and beyond, the adoption of machine learning techniques has surfaced as a notable strategy for battery modeling. Machine learning models are primarily utilized for forecasting the forthcoming state of batteries, with a specific focus on analyzing the State-of-Charge (SOC). Additionally, these models are employed to assess the State-of-Health (SOH) and predict the Remaining Useful Life (RUL) of batteries. Moreover, offering clear explanations for abnormal battery usage behavior is crucial, empowering users with insights needed for informed decision-making, build trust in the system, and ultimately enhance overall satisfaction. This paper presents SOCXAI, a novel algorithm designed for precise estimation of batteries’s SOC. Our proposed model utilizes a Convolutional Neural Network (CNN) architecture to efficiently estimate the twenty five future values of SOC, rather than a single unique value. We also incorporate a SHApley Additive exPlanations (SHAP) based post-hoc explanation method into our method focusing on the current feature values for deeper prediction insights. Furthermore, to detect abnormal battery usage behavior, we employ a 2-dimensional matrix profile-based approach on the time series of current values and their corresponding SHAP values. This methodology facilitates the detection of discords, which indicate irregular patterns in the battery usage. Our extensive empirical evaluation,using diverse real-world benchmarks, demonstrates our approach effectiveness, showcasing its superiority over state-of-the-art algorithms.

ABSTRACT. The rapid evolution of smart manufacturing and the pivotal role of Automated Guided Vehicles (AGVs) in enhancing operational efficiency, underscore the necessity for robust anomaly detection mechanisms. This paper presents a comprehensive approach to detecting anomalies based on AGV telemetry data, leveraging the potential of machine learning (ML) algorithms to analyze complex data streams and time series signals. By focusing on the unique challenges posed by real-world AGV environments, we propose a methodology that integrates data collection, preprocessing, and the application of specific AI/ML models to accurately identify deviations from normal operations. Our approach is validated through extensive experiments on datasets featuring anomalies caused by mechanical wear or excessive friction and issues related to tire and wheel damage, employing LSTM and GRU networks, alongside traditional classifiers like K-Neighbors and SVM. The results demonstrate the efficacy of our method in predicting momentary power consumption as an indicator of mechanical anomalies, and in classifying wheel-related issues with high accuracy. This work not only contributes to the enhancement of predictive maintenance strategies but also provides valuable insights for the development of more resilient and efficient AGV systems in smart manufacturing environments.

ABSTRACT. Augmented reality (AR)-based flight simulation reshapes how pilots are trained, offering an immersive environment where commercial and fighter pilots can be trained at low cost with minimal use of fuel and safety concerns. This study conducts a pioneering comparative analysis of marker-based tracking and SLAM technologies within the Microsoft HoloLens 2 platform, mainly focusing on their efficacy in landing manoeuvre simulations. Our investigation incorporates an experimental setup where marker-based tracking overlays interactive video tutorials onto a simulated cockpit, enhancing the realism and effectiveness of landing procedures. The experiment demonstrates that marker-base systems ensure high precision within 5 cm and 15 cm from the HoloLens 2 camera, proving indispensable for procedural training that requires exact overlay precision. Conversely, the native SLAM algorithm, while lacking the same level of precision, offers flexibility and adaptability by accurately mapping the cockpit and superimposing virtual information in dynamic, markerless conditions. The study juxtaposes these technologies, revealing a trade-off between precision and adaptability, and suggests an integrative approach to leverage their respective strengths. Our findings provide pivotal insights for developers and training institutions to optimize AR flight simulation training, contributing to advanced, immersive pilot training programs.

ABSTRACT. In this talk, we propose an adaptive mesh-refining based on the multi-level algorithm and derive a unified a posteriori error estimate for a class of nonlinear problems in the abstract framework of Brezzi, Rappaz, and Raviart. The multi-level algorithm on adaptive meshes retains quadratic convergence of Newton's method across different mesh levels both theoretically and numerically.

As applications, we consider the pseudostress-velocity formulation of Navier-Stokes equations and the standard Galerkin formulation of semilinear elliptic equations. Several numerical examples are presented to test the performance of the algorithm and validity of the theory developed.

This is joint work with Dongho Kim and Boyoon Seo.

ABSTRACT. Convection-diffusion problems in highly convective flows can exhibit complicated features such as sharp shocks and shear layers which involve steep gradients in their solutions. As a consequence, developing an efficient computational solver to capture these flow features requires the adjustment of the local scale difference between convection and dif- fusion terms in the governing equations. In this study, we propose a monotonicity preserving backward characteristics scheme combined with a second-order BDF2-Petrov-Galerkin finite volume method to deal with the multiphysics nature of the problem. Unlike the conventional Eulerian techniques, the two-step backward differentiation procedure is applied along the characteristic curves to obtain a second-order accuracy. Nu- merical results are presented for several benchmark problems including sediment transport in coastal areas. The obtained results demonstrate the ability of the new algorithm to accurately maintain the shape of the computed solutions in the presence of sharp gradients and shocks.

ABSTRACT. The stability of the initial platelet aggregates is relevant in both hemostasis and thrombosis. Understanding the structural stresses of such aggregates under different flow conditions is crucial to gaining insight into the role of platelet activation and binding in the more complex process of clot formation. In this work, a three-dimensional implicit partitioned fluid-structure interaction (FSI) model is presented to study the deformation and structural stress of platelet aggregates in specific blood flow environments. Platelet aggregates are considered as porous mediums in the model. While the platelets are considered incompressible, the overall structure becomes compressible through its porous nature. The FSI model couples a fluid solver based on Navier-Stokes equations and a porosity-dependent compressible neo-Hookean material to capture the mechanical characteristics of the platelet aggregates. A parametric study is performed to explore the influence of porosity and applied body force on this material. Based on in vitro experimental data, the deformation and associated stress of a low shear aggregate and a high shear aggregate under different flow conditions are evaluated. This FSI framework offers a way to elucidate the complex interaction between blood flow and platelet aggregates and is applicable to a wider range of porous biomaterials in flow.

ABSTRACT. Generalized Hydrodynamic Equations (GHE) are investigated for simulation of turbulent flows. GHE were derived from the Generalized Boltzmann Equation (GBE) by Alexeev (1994). GBE was obtained by first principles from the chain of Bogolubov kinetic equations and considers particles of finite dimensions. The GHE has new terms, temporal and spatial fluctuations, compared to the Navier-Stokes equations (NSE). These new terms have a timescale multiplier tau, and the GHE becomes the NSE when tau=0. The nondimensional tau is a product of the Reynolds number and the squared length scales ratio, tau=Re*(l/L)^2,where l is the apparent Kolmogorov length scale, and L is a hydrodynamic length scale. The BFS flow modeling results by Navier-Stokes equations simulations cannot match the experimental data for Re>450 (Jiang, 1993). Additional equations are required for the turbulence model to be added to the NSE, which typically have two to five parameters to be tuned for specific problems. We show that the GHE does not require an additional turbulence model, whereas the turbulent velocity is in good agreement with the experimental results (Kim 1980).

In this study, the 2D turbulent flow over a BFS with height H=L/3 (where L is the channel height) at Reynolds number Re=132000 was investigated using finite-element solutions of the GHE, and compared to the solutions from the Navier-Stokes equations, k–epsilon turbulence model, and experimental results. The comparison included the velocity profiles, recirculation zone length, and velocity flow field. The obtained data confirm that the GHE results are in good agreement with the experimental results, while other approaches are far from the experimental data.

ABSTRACT. Wearing a face mask is widely acknowledged as a critical defense against the transmission of the novel coronavirus (COVID-19) and influenza. This research focuses on the deformation of face masks during a cough and uses fluid dynamics simulations to more precisely predict the trajectory of virus-laden droplets. By employing motion capture technology, we measured the mask's displacement, which reaches up to 6 mm during a cough. Moreover, this paper delves into how the mask's deformation influences the movement of these droplets. We created a model for a small, spherical droplet and analyzed its dispersion by solving its motion equation, factoring in the cough's flow rate, droplet size distribution, and evaporation, all while considering the mask's deformation. Our findings reveal that mask deformation leads to a 7% reduction in average flow velocity compared to analyses using a non-deforming mask. Additionally, the distance droplets disperse increases over time when mask deformation is considered. As a practical application, we analyzed droplet dispersion in a scenario where a wheelchair is being pushed, utilizing airflow data that accounts for mask deformation. This analysis indicated that pushing a wheelchair at a speed of 0.5 m/s significantly raises the infection risk for individuals behind it.

ABSTRACT. When solving numerically partial differential equations such as the Navier-Stokes equations, higher-order finite difference schemes are occasionally applied for spacial descretization. Compact finite difference schemes are one of the finite difference schemes and can be used to compute the first-order derivative values with smaller number of stencil grid points, however, a linear system of equations with a tridiagonal or pentadiagonal matrix derived from the schemes have to be solved. In this paper, an asymptotic parallel solver for a reduce matrix, that obtained from the Mattor's method in a computation of the first-order derivatives with an eighth-order compact difference scheme under a periodic boundary condition, is proposed. The asymptotic solver can be applied as long as the number of grid points of each Cartesian coordinate in the parallelized subdomain is 64 or more, and its computational cost is lower than that of the Mattor's method. A direct numerical simulation code has also been developed using the two solvers for compressible turbulent flows under isothermal conditions, and optimized on the vector supercomputer SX-Aurora TSUBASA. The optimized code is 1.7 times faster than the original one for a DNS with 2048^3 grid points and the asymptotic solver achieves approximately a 4-fold speedup compared to the Mattor's solver. The code exhibits excellent weak scalability.

ABSTRACT. Message Passing Interface (MPI) is the predominant and most extensively utilized programming model in the High Performance Computing (HPC) area. The standard only provides bindings for the low-level programming languages C, C++, and Fortran. While efforts are being made to offer MPI bindings for other programming languages, the support provided may be limited, potentially resulting in functionality gaps, performance overhead, and compatibility problems. To deal with those issues, we introduce MPI4All, a novel tool aimed at simplifying the process of creating efficient MPI bindings for any programming language. MPI4All is not dependent on the MPI implementation, and adding support for new languages does not require significant effort. The current version of MPI4All includes binding generators for Java and Go programming languages. We demonstrate their good performance with respect to other state-of-the-art approaches.

ABSTRACT. The objective of the paper is to describe computational methods and techniques of investigation of certain algebraic structures needed in order to apply the results in concrete problems in mathematical control theory of nonlinear systems. Contemporary theoretic research requires more and more sophisticated tools for a possible application of the results. In the paper we propose computational tools and techniques for a certain type of simplification of driftless control systems. Such simplification still preserve most crucial properties of the original ones like controllability, but the simplified system have a special feedforward form that is much easier to integrate or allows to solve other problems in control theory. We present the computational procedure and foundations of the library as the extension of existing software libraries in Python language. The approach is illustrated with some numerical experiments and discussion about related computational issues.

ABSTRACT. We propose a technique to circumvent the direct computation of singular surface integrals in parametric integral equation system (PIES) employed for solving three-dimensional potential problems. It is based on the regularization of the original singular PIES formula, resulting in the simultaneous elimination of both strongly and weakly singular integrals. As a result, there is the possibility of numerically calculating the values of all integrals in the obtained formula using standard Gaussian quadrature. The evaluation of accuracy for the proposed approach is examined through an illustrative case, specifically focusing on the steady-state temperature field distribution problem.

ABSTRACT. The paper presents the interval fast parametric integral equations system (IFPIES) applied to solve multi-connected boundary value problems (BVPs) with uncertainly defined complex shapes of a boundary. The method is similar to the fast PIES, which uses the fast multipole method to speed up solving BVPs and reduce RAM utilization. However, modelling uncertainty in the IFPIES uses interval numbers and directed interval arithmetic. Segments created the boundary have the form of the interval B\'ezier curves of the third degree (curvilinear segments) or the first degree (linear segment). The curves also required some modifications connected with applied directed interval arithmetic. In the presented paper, the reliability and efficiency of the IFPIES solutions were verified on multi-connected BVPs with uncertainly defined complex domain shapes. The solutions were compared with the ones obtained by the interval PIES only due to the lack of examples in the literature of solving uncertainly defined BVPs. All presented tests confirm the high efficiency of the IFPIES method.

ABSTRACT. Numerical methods are important tools for design of modern antennas. Trust-region (TR) methods coupled with data-efficient surrogates based on finite differentiation (FD) represent a popular class of antenna design algorithms. However, TR performance is subject to FD setup, which is normally determined a priori based on rules-of-thumb. In this work, the effect of FD perturbations on the performance of TR-based design is evaluated on a case study basis concerning a total of 80 optimizations of a planar antenna structure. The obtained results demonstrate that, for the considered radiator, the performance of the final designs obtained using different FD setups may vary by as much as 18 dB (and by over 4 dB on average). At the same time, the a priori perturbations in a range between 1.5% and 3% (w.r.t. the initial design) seem to be suitable for maintaining (relatively) consistent and high-quality results.

ABSTRACT. We present a new approach for real-time retrieval and classification of solar images using a proposed sector-based image hashing technique. To this end, we generate intermediate hand-crafted features from automatically detected active regions in the form of layer-sector-based descriptors. Additionally, we employ a small fully-connected autoencoder to encode and finally obtain the concise Layer-Sector Solar Hash. By reducing the amount of data required to describe the Sun images, we achieve almost real-time retrieval speed of similar images to the query image. Since solar AIA images are not labeled, for the purposes of the presented test experiments, we consider images produced within a short time frame (typically up to several hours) to be similar. This approach has several potential applications, including searching, classifying, and retrieving solar flares, which are of critical importance for many aspects of life on Earth.

ABSTRACT. Designing deep learning architectures is a challenging and time-consuming task. To overcome this problem, Neural Architecture Search (NAS), which automatically searches for a network topology is often used. While existing NAS methods mainly focus on image classification tasks, particularly 2D medical images, this study presents an evolutionary NAS approach for 2D and 3D Medical image classification. We defined two different search spaces for 2D and 3D datasets and performed a comparative study of different meta-heuristics used in different NAS studies. Moreover, zero-cost proxies have been used to evaluate the performance of deep neural networks, which helps reduce the searching cost of the overall approach. Furthermore, recognizing the importance of Data Augmentation (DA) in model generalization, we propose a genetic algorithm based automatic DA strategy to find the optimal DA policy for a given dataset. Experiments on MedMNIST and BreakHIS datasets demonstrate the effectiveness of our approach, showcasing competitive results compared to existing AutoML approaches.

ABSTRACT. Music is a ubiquitous phenomenon. In today’s world, no one can imagine life without its presence, and no one questions its significance in human life. This is not a new phenomenon but has been prevalent for hundreds of years. Therefore, an automated approach to understanding music plays a nontrivial role in science. One of the many tasks in Music Information Retrieval is the categorization of musical compositions. In this paper, the authors address the rarely explored topic of classifying traditional musical compositions from different cultures into regions (continents), subregions, and countries. A newly created dataset is presented, along with preliminary classification results using well-known classifiers. The presented work marks the beginning of a long and fascinating scientific journey.

ABSTRACT. The Domain Name System (DNS) is a crucial infrastructure of the Internet, yet it is also a primary medium for disseminating illicit content. Researchers have proposed numerous methods to detect malicious domains, with association-based approaches achieving relatively good performance. However, these methods encounter limitations in detecting malicious domains within isolated nodes and heavily relying on labeled data to improve performance. In this paper, we propose a semi-supervised malicious domain detection model named SemiDom, which is based on meta pseudo labeling. Firstly, we use associations among DNS entities to construct a semantically enriched domain association graph. In particular, we retain isolated nodes within the dataset that lack relationships with other entities. Secondly, a teacher network computes pseudo labels on the unlabeled nodes, which effectively augments the scarce labeled data. A student network utilizes these pseudo labels to transform both the structure and attribute features to domain labels. Finally, the teacher network is constantly optimized based on the student's performance feedback on the labeled nodes, enabling the generation of more precise pseudo labels. Extensive experiments on the real-world DNS dataset demonstrate that our proposed method outperforms the state-of-the-art methods.

ABSTRACT. Abstract. Over recent years, archaeologists have started to use object detection methods in satellite images to search for potential archaeological sites. Within image object recognition, due to its ability to recognize objects with great accuracy, convolutional neural networks (CNN) are becoming increasingly popular. This study compares the performance of existing deep learning algorithms for detecting small megalithic monuments in satellite image-ry. A satellite image dataset is used for object detection using convolutional neural networks, namely RCNN (Region-based Convolutional Neural Net-works) and YOLO (You Only Look Once). Using this dataset and after ade-quate preprocessing, results showed that this is a feasible approach for archeological image prospection, with RCNN achieving a precision of 93% in small monuments’ detection.

ABSTRACT. Physics-informed Neural Networks (PINNs), a class of neural models that are trained by minimizing a combination of the residual of the governing partial differential equation and the initial and boundary data, have gained immense popularity in the natural and engineering sciences. Despite their observed empirical success, an analysis of the training efficiency of residual-driven PINNs at different architecture depths is poorly documented. Usually, neural models used for machine learning tasks such as computer vision and natural language processing have deep architectures, that is, a larger number of hidden layers. In PINNs, we show that for a given trainable parameter count (model size), a shallow network (less layers) converges faster than a deep network (more layers) for the same error characteristics. To illustrate this, we examine the one-dimensional Poisson's equation and evaluate the gradient for residual and boundary loss terms. We show that the characteristics of the gradient of the loss function are such that for residual loss, shallow architectures converge faster. Empirically, we show the implications of our theory through various experiments.

ABSTRACT. The Deep Fourier Residual (DFR) method, a subset of Variational Physics-Informed Neural Networks (VPINN), aims to solve PDEs using neural networks. The loss function in VPINN approximates the dual norm of the PDE’s weak-residual, ensuring that minimizing the loss corresponds to reducing the error in the solution. While previous methodologies like Robust Variational PINNs face challenges with costly matrix inversions, DFR leverages orthonormal basis functions for efficiency. DFR extends to function spaces like H1(Ω) and H(curl, Ω) by utilizing spectral representations of dual norms. This talk introduces an extension of DFR to include adaptive strategies on polygonal domains, refining meshes for improved accuracy, particularly evident in singular cases like the L-shape problem. Numerical results demonstrate the effectiveness of the proposed approach.

ABSTRACT. This research presents an efficient computational method for retrofitting buildings by employing an active learning-based ensemble machine learning (AL-Ensemble ML) approach developed in OpenSees, Python and MATLAB. The results of the study shows that the AL-Ensemble ML model provides the most accurate estimations of interstory drift (ID) and residual interstory drift (RID) for steel structures using a dataset of 2-, to 9-story steel structures considering four soil type effects. To prepare the dataset, 3584 incremental dynamic analysis (IDA) were performed on 64 structures. The re-search employs 6-, and 8-story structures to validate the AL-Ensemble ML model's effectiveness, showing it achieves the highest accuracy among conventional ML models, with an R2 of 98.4%. Specifically, it accurately predicts the RID of floor levels in a 6-story structure with an accuracy exceeding 96.6%. Additionally, the programming code identifies the specific dam-aged floor level in a building, facilitating targeted local retrofitting instead of retrofitting the entire structure promising a reduction in computational time and retrofitting costs while enhancing prediction accuracy.

ABSTRACT. Multivariate time series anomaly detection is a crucial prob- lem in many industrial and research applications. Timely detection of anomalies allows, for instance, to prevent defects in manufacturing pro- cesses and failures in cyberphysical systems. Deep learning methods are preferred among others for their accuracy and robustness for the analysis of complex multivariate data. However, a key aspect is being able to extract predictions in a timely manner, to accommodate real-time require- ments in different applications. In the case of deep learning models, model reduction is extremely important to achieve optimal results in real-time systems with limited time and memory constraints. In this paper, we address this issue by proposing a novel compression method for deep autoencoders that involves three key factors. First, pruning reduces the number of weights, while preventing catastrophic drops in accuracy by means of a fast search process that identifies high sparsity levels. Second, linear and non-linear quantization reduces model complexity by reducing the number of bits for every single weight. The combined contribution of these three aspects allow the model size to be reduced, by removing a sub- set of the weights (pruning), and decreasing their bit-width (quantization). As a result, the compressed model is faster and easier to adopt in highly constrained hardware environments. Experiments performed on popu- lar multivariate anomaly detection benchmarks, show that our method is capable of achieving significant model compression ratio (between 80% and 95%) without a significant reduction in the anomaly detection performance

ABSTRACT. The following work presents a new method to solve Poisson's equation and, more generally, sparse linear systems using graph neural networks. We propose a supervised approach to solve the discretized representation of Poisson's equation at every time step of a simulation. This new method will be applied to plasma physics simulations for Hall-Effect Thruster's modeling, where the gradient of the electric potential must be computed to get the electric field, necessary to model the plasma's behavior. In this setting, solving Poisson's equation using classical iterative methods represents a major part of the computational costs. This is even more critical for unstructured meshes which increases the problem's complexity. To accelerate the computational process, we propose a graph neural network to give an initial guess of Poisson's equation solution. The new method introduced in this article has been designed to handle any form of meshing structure, including structured and unstructured grids, as well as any kind of sparse linear systems. Once trained, the neural network would be used inside a numerical simulation in inference to give an initial guess of the solution for each time step of a simulation for all right-hand sides of the linear system and all previous time step solutions. In most industrial cases, Hall-Effect thrusters' modeling requires a large unstructured mesh that cannot be held by one single processor in terms of memory capacity. We then propose a partitioning strategy to tackle the challenge of solving linear systems on large unstructured grids when they cannot be on a single processor.

ABSTRACT. In this paper, adaptive hyperparameter optimization (HPO) strategies within the efficient global optimization (EGO) with neural network (NN)-based prediction and uncertainty (EGONN) algorithm are proposed. These strategies utilize Bayesian optimization and multi-armed bandit optimization to tune HPs during the sequential sampling process either every iteration (HPO-1itr) or every five iterations (HPO-5itr). Through experiments using the three-dimensional Hartmann function and evaluating both full and partial sets of HPs, adaptive HPOs are compared to traditional static HPO (HPO-static) that keep HPs constant. The results reveal that adaptive HPO strategies outperform HPO-static, and the frequency of tuning and number of tuning HPs impact both the optimization accuracy and computational efficiency. Specifically, adaptive HPOs demonstrate rapid convergence rates (HPO-1itr at 28 iterations, HPO-5itr at 26 for full HPs; HPO-1itr at 13, HPO-5itr at 28 iterations for selected HPs), while HPO-static fails to approximate the minimum within the allocated 45 iterations for both scenarios. Mainly, HPO-5itr is the most balanced approach, found to require 21% of the time taken by HPO-1itr for tuning full HPs and 29% for tuning a subset of HPs. This work demonstrates the importance of adaptive HPO and sets the stage for future research.

ABSTRACT. Almost all optimization algorithms have algorithm-dependent parameters, and the setting of such parameter values can largely influence the algorithm's behavior under consideration. Thus, proper parameter tuning should be carried out to ensure that the algorithm used for optimization performs well and is sufficiently robust for solving different types of optimization problems. In this study, the Firefly Algorithm (FA) is used to evaluate the possible influence of its parameter values on its efficiency. Parameter values are randomly initialized using the standard Monte Carlo and Quasi Monte-Carlo methods. The values are then used for tuning the FA. Two benchmark functions and a spring design problem are used to test the robustness of the tuned FA. From the preliminary findings, it can be deduced that both the Monte Carlo method and Quasi-Monte Carlo method produces similar results in terms of optimal fitness values. Numerical experiment using the two different methods on both benchmark functions and the spring design problem showed no significant variations in the final fitness values, irrespective of the various sample values selected during the simulations. This insensitivity indicates the robustness of the FA.

ABSTRACT. Non-intrusive reduced order modeling methods (ROMs) have become increasingly popular for science and engineering applications such as predicting the field-based solutions for aerodynamic flows. A large sample size is, however, required to train the models for global accuracy. In this paper, a novel adaptive sampling strategy is introduced for these models that uses field-based uncertainty as a sampling metric. The strategy uses Monte Carlo simulations to propagate the uncertainty in the prediction of the latent space of the ROM obtained using a multi-task Gaussian process to the high-dimensional solution of the ROM. The high-dimensional uncertainty is used to discover new sampling locations to improve the global accuracy of the ROM with fewer samples. The performance of the proposed method is demonstrated on the environment model function and compared to one-shot sampling strategies. The results indicate that the proposed adaptive sampling strategies can reduce the mean relative error of the ROM to the order of 0.0008 which is a 20% and 27% improvement over the Latin hypercube and Halton sequence sampling strategies, respectively at the same number of samples.

ABSTRACT. Solving an optimal control problem consists in finding a control structure and corresponding switching times. Unlike in a bang-bang case, switching to a singular control perturbs the control structure. The perturbation of one of the switching times affects any subsequent singular intervals in the control, as the trajectories move along different singular arcs with different values of singular controls. It makes the problem of finding optimal solutions extremely difficult. In this paper, we discuss a gradient method for solving optimal control problems, when singular intervals are present in the optimal structure. The method is based on applying the necessary conditions of optimality given by the Pontryagin Maximum Principle, where the control variable enters the Hamiltonian linearly. To demonstrate the method, we formulate a~nonlinear optimal control problem and then, using the proposed algorithm, we solve the problem and find structures of optimal controls and corresponding switching times. Lastly, we compare the results with results obtained using three popular optimisation modelling languages: Pyomo, AMPL and JuMP. These languages serve as interfaces for solving the optimal control problem with the non-linear optimisation algorithm Ipopt. Our case study shows that the presented method not only computes the switching times accurately, but also moves precisely along the singular arc.

ABSTRACT. A classical non-stationary (n+1)-body planetary problem with n bodies of variable mass moving around the central star on quasi-elliptic orbits is considered. In addition to the mutual gravitational attraction, the bodies may be acted on by reactive forces arising due to anisotropic variation of their masses. The problem is analyzed in the framework of Newtonian's formalism and the differential equations of motion are derived in terms of the osculating elements of aperiodic motion on quasi-conic sections. These equations can be solved numerically and their solution will describe the motion of the bodies in detail. However, due to the orbital motion of the bodies the perturbing forces include many terms describing short-period oscillations. Therefore, to obtain the solution with high precision one needs to choose very small step size or to use an adoptive step size method and this increase a time of calculation substantially. As we are interested in the long-term behaviour of the system it will be necessary to perform additional calculations in order to extract a secular part of the solution. To simplify the calculations we expand the perturbing forces into power series in terms of eccentricities and inclinations which are assumed to be small and average these equations over the mean longitudes of the bodies. Finally, we obtain the differential equations describing the evolution of orbital parameters over a long period of time. As an application, we have solved the evolution equations numerically in the case of n=3 and demonstrated an influence of the mass variation on the motion of the bodies. All the relevant symbolic and numeric calculations are performed with the aid of the computer algebra system Wolfram Mathematica.

ABSTRACT. In this contribution, we developed a software tool for collecting information on the data traffic via control plane of an operating optical network. From this data, demand matrix elements were calculated and used to numerically estimate the edge occupancy in the optical network studied. For this purpose, a detailed network model was formulated with cost function and constraints. The formulated network model leads to an optimization problem, which was efficiently solved by meta-heuristic algorithms. Finally, statistical methods were used to model forecasting, in terms of the probability of the edge occupancy, under a Markov process approximation. Additionally, on the basis of the numerical results obtained, the scalability of the applied heuristic and statistical methods was analyzed.

ABSTRACT. Chronic Kidney Disease (CKD) is a common disease with high incidence and high risk for the patients' health when it degrades to its most advanced stages. When detected early, it is possible to slow down the progression of the disease, leading to an increased survival rate and lighter treatment. As a consequence, many prediction models have emerged for the prediction of CKD. However, few of them manage to efficiently predict the onset of the disease months to years prior. In this paper, we propose an artificial neural network combining the strengths of convolution and involution layers in order to predict the degradation of CKD to its later stages, based on a set of 25 common laboratory analyses as well as the age and gender of the patient. Using a dataset from a French medical laboratory containing more than 400 000 patients, we show that our model achieves better performance than state-of-the-art models, with a recall of 83%, F1-score of 76%, and 97% overall accuracy. The proposed method is flexible and easily applicable to other diseases, offering encouraging perspectives in the field of early disease prediction, as well as the use of involution layers for deep learning with time series.

ABSTRACT. This work introduces an enhanced methodology in the domain of eXplainable Artificial Intelligence (XAI) for visualizing local explanations of black-box, feature-based models, such as LIME and SHAP, enabling both domain experts and non-specialists to identify the segments of Time Series (TS) data that are significant for machine learning model interpretations across classes. By applying this methodology to electrocardiogram (ECG) data for anomaly detection, distinguishing between healthy and abnormal segments, we demonstrate its applicability not only in healthcare diagnostics but also in predictive maintenance scenarios. Central to our contribution is the development of the AUC Perturbational Accuracy Loss metric (AUC-PALM), which facilitates the comparison of explainer fidelity across different models. We advance the field by evaluating various perturbation methods, demonstrating that perturbations centered on time series prototypes and those proportional to feature importance outperform others by offering a more distinct comparison of explainer fidelity with the underlying black-box model. This work lays the groundwork for broader application and understanding of XAI in critical decision-making processes.

ABSTRACT. This study investigates the impact of advanced computational methodologies on brain tumor segmentation in medical imaging, addressing challenges like interobserver variability and biases. The DeepLabV3plus model with ResNet50 integration is rigorously examined and augmented by diverse image enhancement techniques. The hybrid CLAHE-HE approach achieves exceptional efficacy with an accuracy of 0.9993, a Dice coefficient of 0.9690, and a Jaccard index of 0.9404. Comparative analyses against established models, including SA-GA, Edge U-Net, LinkNet, MAG-Net, SegNet, and Multi-class CNN, consistently demonstrate the proposed method’s robustness. The study underscores the critical need for continuous research and development to tackle inherent challenges in brain tumor segmentation, ensuring insights translate into practical applications for optimized patient care. These findings offer substantial value to the medical imaging community, emphasizing the indispensability of advancements in brain tumor segmentation methodologies. The study outlines a path for future exploration, endorsing ensemble models like U-Net, ResNet-U-Net, VGG-U-Net, and others to propel the field toward unprecedented frontiers in brain tumor segmentation research.

ABSTRACT. Venous thromboembolism (VTE) poses a significant risk to patients undergoing cancer treatment, particularly in the context of advanced and metastatic disease. In the realm of neuro-oncology, the incidence of VTE varies depending on tumor location and stage, with certain primary and secondary brain tumors exhibiting a higher propensity for thrombotic events. In this study, we employ advanced machine learning techniques, specifically XGBoost, to develop identifying models for predictors searching associated with VTE risk in patients with gliomas. By comparing the diagnosis testing accuracy of our XGBoost models with traditional logistic regression approaches, we aim to enhance our understanding of VTE prediction in this population. Our findings contribute to the growing body of literature on thrombosis risk assessment in cancer patients and may inform the development of personalized prevention and treatment strategies to mitigate the burden of VTE in individuals with gliomas at the hospital term.

ABSTRACT. Cervical cancer is the fourth most common cancer among women. Every year, more than 200,000 women die due to cervical cancer; however, it is one of the preventable diseases if detected early. The purpose of this study is to detect cervical cancer by identifying the cytoplasm and nuclei of the background using deep learning techniques to automate the separation of a single cell. To preprocess the image, resizing and enhancement are adopted by adjusting the brightness and contrast of the image to remove noise in the image. For model training, we divide the data into 80% training and 20% testing. To segment the image, Unet is the original network and serves as a baseline, with VGG19, ResNet50, MobileNet, EfficientNetB2 and Dense-Net121 as the backbone. In cytoplasmic segmentation, EfficientNetB2 achieves a precision of 99.02%, while DenseNet121 achieves an accuracy of 98.59% for a single smear cell. In nuclei segmentation, EfficientNetB2 achieves an accuracy of 99.86%, surpassing ResNet50, which achieves 99.85%. As a result, deep learning-based image segmentation produces promising results in separating the cytoplasm from the background to detect cervical cancer. This is helpful for cytotechnicians in diagnosis and decision-making.

ABSTRACT. Intrusion detection systems (IDS) play a critical role in protecting networks from various cyber threats. Currently, AI-based intrusion detection methods stand as the mainstream, yet they grapple with the challenges of interpretability and high computational costs. Conversely, rule-based approaches offer ease of comprehension and lower computational overhead, but their development demands extensive expertise. This paper proposes an intelligent framework for generating intrusion detection rules, which integrates the strong representational capabilities of AI detectors while retaining the advantages of rule-based detection. Initially, the framework involves training a TextCNN model for malicious traffic payload classification. The parameters of this model, along with the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm, are employed to analyze critical fields in captured traffic payloads. Subsequently, a comprehensive list of keywords is obtained through a sensitive words aggregation algorithm, and regular expressions are generated to describe the detection content. These regular expressions undergo fine-tuning to reduce their false positive rate. Furthermore, adhering to the syntax of Suricata rules, they are formulated into intrusion detection rules. The proposed method is evaluated on two publicly available datasets, with experimental results demonstrating commendable detection efficacy for the generated intrusion detection rules.

ABSTRACT. With the rapid development of AI-based technology, social bot detection is becoming an increasingly challenging task to combat the spread of misin-formation and protect the authenticity of online resources. Existing graph-based social bot detection approaches primarily rely on the topological structure of the Twittersphere but often overlook the diverse influence dynamics across different relationships. Moreover, these methods typically aggregate only direct neighbors based on transitive learning, limiting their effectiveness in capturing the nuanced interactions within evolving social Twittersphere. In this paper, we propose BotRGA, a novel Twitter bot detection framework based inductive representation learning. Our method begins with extracting the semantic features from Twitter user profiles, descriptions, tweets and constructing a heterogeneous graph, where nodes represent users and edges represent relationships. We then propose a relational graph aggregation method to learn node representations by sampling and aggregating the features from both direct and indirect neighbors. Additionally, we evaluate the importance of different relations and fuse the node’s representations across diversified relations with semantic fusion networks. Finally, we classify Twitter users into bots or genuine users and learn model parameters. Extensive experiments conducted on two comprehensive Twitter bot detection benchmarks demonstrate that the superior performance of BotRGA compared to state-of-the-art methods. Additional studies also confirm that the effectiveness of our proposed relational graph aggregation, semantic fusion networks, and strong generalization ability to new and previously unseen user communities.

ABSTRACT. This paper aims to explore the problem of air pollution forecasting, especially the particulate matter (PM) concentration in the air. Other quantities such as air temperature, atmospheric pressure, and relative humidity are also considered. Moreover, a large part of the discussion in this paper can be extended and applied to a variety of other quantities which are stored and expressed as data series. The goal is to evaluate different time series forecasting models on a selected air pollution data set. The proposed model is compared with other implemented state-of-the-art methods in order to validate whether it could be a reliable pick for air pollution forecasting problem.

ABSTRACT. The concept of the Internet of Things (IoT) has been widely used in many applications. IoT devices can be exposed to various external factors, such as network congestion, signal interference, and limited network bandwidth. This paper proposes an open-source uChaos software tool for the ZephyrOS real-time operating system for embedded devices intended to inject failures in a controlled manner to improve their error handling algorithms. The proposed novel framework fills the gap in the chaos engineering tools for the cloud-edge continuum. In the paper, we also discuss the typical failures of IoT devices and the potential use cases of the solution.

ABSTRACT. Pore-scale simulation is widely used in two-phase flow to study capillary pressure as a function of saturation in porous media. However, two-phase flow is still challenging when fluids are binary mixtures whose species can be dissolved into two phases. In this work, the pore-network model coupling phase equilibrium calculations is employed to investigate the influence of binary mixtures on capillary pressure. Our results show that interfacial tension of binary mixtures is the key parameter to affect capillary pressure-saturation curves. Phase equilibrium calculations show that interfacial tension in binary mixtures can be formulated as a function of fluid pressure at the constant temperature. When imposing the constant capillary pressure, variations of saturation can be related to fluid pressure for elaborating the thermodynamics of binary mixtures.

ABSTRACT. The algorithm presented in this study uses a quasi-gasdynamic approach to address unsteady, compressible flows over a wide range of Mach numbers. This implementation is carried out within the AMReX open platform, which uses adaptive mesh refinement technology and facilitates the parallelization of computations on GPU architectures. To validate its effectiveness, the solver was applied to a specific scenario involving the problem of a composite vortex-stationary shock wave interaction, with flow parameter values of Mv = 0.9 and Ms = 1.5. To assess its performance, cross-validation was performed with OpenFOAM-based solvers, specifically rhoCentralFoam and QGDFoam. Schlieren fields are used to evaluate the oscillations of the numerical schemes and algorithms, while the resolution capabilities of the algorithm are assessed by comparing the density fields in five cross-sections with reference values.

ABSTRACT. For the Stokes problem with pressure Dirichlet boundary conditions, we propose an Enriched Mini element. For both the Mini element and the Enriched Mini element, we show that they are inf-sup stable. Unexpectedly, they yield wrong convergent finite element solutions for the singular velocity solution. On the contrary, the Taylor-Hood element, which is still inf-sup stable, gives correct convergence. However, how to analyze the convergence becomes open.

ABSTRACT. Unstructured flux-limiters convective schemes designed in the framework of the unstructured conservative level-set (UCLS) method, are assessed for transport phenomena in two-phase flows. Transport equations are discretized by the finite volume method on 3D collocated unstructured meshes. The central difference scheme discretizes the diffusive term. Gradients are evaluated by the weighted least-squares method. The fractional-step projection method solves the pressure-velocity coupling in momentum equations. Numerical findings about the effect of flux limiters on the control of numerical diffusion and improvement of numerical stability in DNS of two-phase flows are reported.

ABSTRACT. Hydrogen is a kind of clean energy with zero carbon emission after combustion, the development of hydrogen energy is one of important ways to solve the greenhouse effect. It is an important means to achieve large-scale utilization of hydrogen by transporting hydrogen to the user terminal through urban gas polyethylene (PE) pipelines. However, due to the property of PE materials, there is a certain amount of hydrogen permeation during transportation, resulting in energy waste and safety issue. In this paper, the solubility coefficient of hydrogen in PE is studied by the Grand canonical Monte Carlo (GCMC) method, and the diffusion coefficient of hydrogen in PE is investigated by the non-equilibrium molecular dynamics (MD) simulation which can consider the real pressure difference inside and outside the PE pipelines. Finally, the permeation coefficient of hydrogen in PE is obtained and investigated. Results show that the solubility coefficient of hydrogen in PE in-creases with the increasing temperature, and the diffusion and permeation coefficients rise with the increase of temperature and pressure difference. In addition, the relationship between above three coefficients with temperature can be described by the Arrhenius law. This study can provide guidance for safe transportation of hydrogen in PE pipelines.