ABSTRACT. Anders will present his organisation, the European High Performance Computing Joint Undertaking (EuroHPC JU). The EuroHPC JU joins together the resources of the European Union, 33 European countries and 3 private partners to develop a World Class Supercomputing Ecosystem in Europe. Anders will present the operational EuroHPC supercomputers located across Europe and give details about their access policy. Free access is already being provided to European research organisations and wider access is planned for the future. Anders will then present some of the JU’s missions, such as the acquisition of new supercomputers including exascale systems and quantum computers, the implementation of an ambitious research and innovation programme supporting European technological and digital autonomy and developing green technologies in HPC, the further strengthening of Europe’s leading position in HPC applications, and other initiatives, part of the effort of the JU to broaden the use of HPC in Europe.

ABSTRACT. Graph machine learning models have been successfully deployed in a variety of application areas. One of the most prominent types of models - Graph Neural Networks (GNNs) - provides an elegant way of extracting expressive node-level representation vectors, which can be used to solve node-related problems, such as classifying users in a social network. However, many tasks require representations at the level of the whole graph, e.g., molecular applications. In order to convert node-level representations into a graph-level vector, a so-called readout function must be applied. In this work, we study existing readout methods, including simple non-trainable ones, as well as complex, parametrized models. We introduce a concept of ensemble-based readout functions that combine either representations or predictions. Our experiments show that such ensembles allow for better performance than simple single readouts or similar performance as the complex, parametrized ones, but at a fraction of the model complexity.

ABSTRACT. The successful application of computational models presupposes access to accurate, relevant and representative datasets. The growth of public data, and the increasing practice of data sharing and reuse, emphasises the importance of data provenance and increases the need for modellers to understand how data processing decisions might impact model output. One key step in the data processing pipeline is that of data integration and entity resolution, where entities are matched across disparate datasets. In this paper, we present a new formulation of data integration in complex networks, and define an approach for understanding how different data integration setups can impact the results of network diffusion models under uncertainty. This approach allows one to systematically characterise potential model outputs in order to create an output distribution that provides a fuller picture.

ABSTRACT. Learning representations of nodes has been a crucial area of the graph machine learning research area. A well-defined node embedding model should reflect both node features and the graph structure in the final embedding. In the case of dynamic graphs, this problem becomes even more complex as both features and structure may change over time. The embeddings of particular nodes should remain comparable during the evolution of the graph, what can be achieved by applying an alignment procedure. This step was often applied in existing works after the node embedding was already computed. In this paper, we introduce a framework - RAFEN - that allows to enrich any existing node embedding method using the aforementioned alignment term and learning aligned node embedding during training time. We propose several variants of our framework and demonstrate its performance on six real-world datasets. RAFEN achieves on-par or better performance than existing approaches without requiring additional processing steps.

ABSTRACT. In this paper, we consider a model wave equation. We perform a sequence of numerical experiments with Physics Informed Neural Network, considering different activation functions, and different ways of enforcing the initial and boundary conditions. We show the convergence of the method and the resulting numerical accuracy for different setups. We show that, indeed, the PINN methodology can solve the problem efficiently and accurately the wave-equations without actually solving a system of linear equations as it happens in traditional numerical methods like, e.g., finite element or finite difference method.In particular, we compare the influence of selected activation functions on the convergence of the PINN method.Our PINN code is available on github: https://github.com/pmaczuga/pinn-comparison/tree/iccs.

ABSTRACT. We are solving the topic clustering problem, where we need to categorize papers with initially available subjects into more consistent and higher-level topics. We approach the task from two perspectives, one is the traditional network science, where we perform community detection on a subject network with the use of the Combo algorithm, and the second is the transformer-based top2vec algorithm which uses sentence-transformer to embed the content of the papers. The comparison between the two approaches was conducted using a dataset of scientific papers on computer science and mathematics collected from the SCOPUS database, and different coherence scores were used as a measure of performance. The results showed that the community detection Combo algorithm was able to achieve a similar coherence score to the transformer-based top2vec. The findings suggest that community detection may be a viable alternative for topic clustering when one has predefined topics, especially when a high coherence score and fast processing time are desired. The paper also discusses the potential advantages and limitations of using Combo for topic clustering and the potential for future work in this area. 

ABSTRACT. The term 'privacy paradox' refers to the apparent inconsistency between privacy concerns and actual behaviour that may also often lead to the dominance of privacy careless over privacy respecting platforms. One of the most important explanations for this phenomenon is based on the effect of social norm on the decisions of individuals to accept or reject a specific platform. However the interdependencies between social norm dynamics and platform adoption have received little attention so far. To overcome this limitation, this article presents a system dynamics simulation model that considers the concept of social norm, shaped by users with diverse privacy concerns, during the adoption process of two alternative social media platforms and identifies the types of situations in which the privacy paradox emerges. The results show a bidirectional minority rule, where (1) the least concerned minority can hinder the more concerned majority from discarding a privacy careless platform but also (2) the most concerned minority can induce the less concerned majority to adopt a privacy respecting platform. Both (1) and, to a lesser extent, (2) are types of situations that reflect the privacy paradox.

ABSTRACT. Today, android dominates the smartphone operating systems market. As per Google, there are over 3 million active android users. With such a large population depending on the platform for their daily activities, a strong need exists to protect android from adversaries. Historically, techniques like signature and behavior were used in malware detectors. However, machine learning and deep learning models have now started becoming a core part of next-generation android malware detectors. In this paper, we step into malware developers/adversary shoes and ask: Are machine learning based android detectors resilient to reinforcement learning based adversarial attacks? Therefore, we propose the RL-MAGE framework to investigate the adversarial robustness of android malware detectors. The RL-MAGE framework assumes the grey-box scenario and aims to improve the adversarial robustness of malware detectors. We designed three reinforcement learning based evasion attacks A2C-MEA, TRPO-MEA, and PPO-MEA, against malware detectors. We investigated the robustness of 30 malware detection models based on 2 features (android permission and intent) and 15 distinct classifiers from 4 different families (machine learning classifiers, bagging based classifiers, boosting based classifiers, and deep learning classifiers). The designed evasion attacks generate adversarial applications by adding perturbations into the malware so that they force misclassifications and can evade malware detectors. The attack agent ensures that the adversarial applications' structural, syntactical, and behavioral integrity is preserved, and the attack's cost is minimized by adding minimum perturbations. The proposed TRPO-MEA evasion attack achieved a mean evasion rate of 93.27% while reducing the mean accuracy of 30 malware detectors from 85.81% to 50.29%. We also propose the ARShield defense strategy to improve the malware detectors' classification performance and robustness. The TRPO-MEA ARShield models achieved 4.10% higher mean accuracy and reduced the mean evasion rate of re-attack from 93.27% to 1.05%. Finally, we conclude that the RL-MAGE framework improved the detection performance and adversarial robustness of malware detectors.

ABSTRACT. The concept of extended cloud requires efficient network infrastructure to support ecosystems reaching form the edge to the cloud(s). Standard approaches to network load balancing deliver static solutions that are insufficient for the extended clouds, where network loads change often. To address this issue, a genetic algorithm based load optimizer is proposed and implemented. Next, its performance is experimentally evaluated and it is shown that it outperforms other existing solutions.

ABSTRACT. Heterogeneous graph neural networks (HGNNs) have shown superior capabilities on graphs that contain multiple types of entities with rich semantic information. However, they are usually (semi-)supervised learning methods which rely on costly task-specific labeled data. Due to the problem of label sparsity on heterogeneous graphs, the performance of these methods is limited, prompting the emergence of some self-supervised learning methods. However, most of self-supervised methods aggregate meta-path based neighbors without considering implicit neighbors that also contain rich information, and the mining of implicit neighbors is accompanied by the problem of introducing irrelevant nodes. Therefore, in this paper we propose a self-supervised deep heterogeneous graph neural networks with contrastive learning (DHG-CL) which not only preserves the information of implicitly valuable neighbors but also further enhances the distinguishability of node representations. Specifically, (1) we design a cross-layer semantic encoder to incorporate information from different high-order neighbors through message passing across layers; and then (2) we design a graph-based contrastive learning task to distinguish semantically dissimilar nodes, further obtaining discriminative node representations. Extensive experiments conducted on a variety of real-world heterogeneous graphs show that our proposed DHG-CL outperforms the state-of-the-arts.

ABSTRACT. This work presents several self-optimization strategies to improve the performance of task-based linear algebra software on heterogeneous systems. The study focuses on Chameleon, a task-based dense linear algebra software whose routines are computed using a tile-based algorithmic scheme and executed in the available computing resources of the system using a scheduler which dynamically handles data dependencies among the basic computational kernels of each linear algebra routine. The proposed strategies are applied to select the best values for the parameters that affect the performance of the routines, such as the tile size or the scheduling policy, among others. Also, parallel optimized implementations provided by existing linear algebra libraries, such as Intel MKL (on multicore CPU) or cuBLAS (on GPU) are used to execute each of the computational kernels of the routines. Results obtained on a heterogeneous system composed of several multicore and multiGPU are satisfactory, with performances close to the experimental optimum.

ABSTRACT. This is a very brief review of the submitted two-page extended abstract:

In the last few years, multiple works have emerged to solve Partial Differential Equations using Neural Networks. For example, the Deep Ritz and Deep Least-Squares methods adapt the classical Ritz and Least-Squares methods to the neural network framework. In this work, we introduce these deep-learning-based methods into a goal-oriented context and carry out theoretical and empirical comparative studies between methods.

ABSTRACT. Many geophysical applications seek to characterize the material property distribution of the Earth’s subsurface. These characterizations are generally accomplished by acquiring measurements and subsequently inverting them using advanced numerical methods. Deep Neural Networks (DNNs) have become increasingly used for this purpose. However, to approximate the inverse operator, it is necessary to obtain the solution to hundreds of thousands of direct problems to construct the training dataset. These problems often require fine or complex finite element grids to describe them accurately since these solutions may involve singularities or intricate geometries. Consequently, managing the computational cost and designing efficient tools that provide accurate solutions for the direct problem becomes essential.

The present work proposes to solve parametric PDEs by efficiently producing hp-adaptive meshes that can address complex direct problems cost-effectively. Specifically, given a problem, we construct a single light-weighted $hp$-adaptive mesh for multiple parameters (such as wavenumbers, material conductivities, etc.). The critical point here is defining the mesh-refinement error indicator, which combines features of several finite element solutions simultaneously. The result is a single $hp$-adapted mesh that delivers fast solutions of sufficient accuracy for any set of parameters within a given range. Thus, we enable the solution of many finite element problems required to train a DNN at a reduced computational cost.

ABSTRACT. Deep learning (DL) strategies have recently emerged as powerful numerical techniques for simulating and inverting problems in computational mechanics and geophysics. For instance, training a Deep Neural Networks (DNNs) enable fast real-time solutions to inverse problems.

Despite the benefits, DL techniques rely on customized design decisions specific to a problem and the available data to produce reliable and robust results. The latter constitutes one of the main difficulties since we typically need enough data that relates sufficient variations of problems' model parameters with the solution of the forward problem, which often corresponds to a Partial Differential Equation (PDE).

In this work, we analyze the influence of the loss functions and study the effects of the quality and quantity of the data in DL inversions. We employ synthetic measurements associated with problems of geophysical interest, and we produce this data with a massive data generator based on a Goal-Oriented $hp$-adaptive Finite Element Method (FEM). Our preliminary results suggest that the proposed approach can provide valuable insight when solving --with DNN techniques-- inverse problems that include data.

ABSTRACT. In this presentation, we will present the Physics-informed Spectral Learning (PiSL) based on a discrete L2 projection to solve the discrete Hodge–Helmholtz decomposition from sparse data. The proposed computational framework combines supervised and unsupervised learning techniques.

Our PiSL computational framework provides spectral (exponential) convergence. Moreover, the method allows us to adaptively build a sparse set of Fourier basis functions with corresponding coefficients by solving a sequence of minimization problems where the set of basis functions is augmented greedily at each optimization problem. Furthermore, the divergence- and curl-free constraints become a finite set of linear algebraic equations in the Fourier structure. We provide several examples to show the performance of our method.

ABSTRACT. In this talk, we propose a Deep Fourier Residual method (DFR), a method of approximating the dual norm of weak-formulation PDE residuals to be used as loss functions for training NNs. The method is based on a spectral representation of the dual norms of the test function space that can be implemented via the Fast Fourier Transform. We will consider applications to problems with H1 test function spaces, corresponding to elliptic PDEs, and H(curl) test function spaces, corresponding to the time-harmonic Maxwell’s equations. Our results show strong correlations between the loss and the error, even in the presence of low-regularity solutions with interface conditions. This demonstrates an advantage over techniques such as Physics Informed Neural Networks (PINNs) and L2-orthogonality-based Variational PINNs (VPINNs), which we show produce undesirable loss-error relationships even in the case of smooth problems.

ABSTRACT. Design of modern antenna systems heavily relies on numerical optimization methods. Their primary purpose is performance improvement by tuning of geometry and material parameters of the structure at hand. For reliability, the process has to be conducted using full-wave electromagnetic (EM) simulation models, which is associated with high computational costs. The problem is aggravated in the case of global optimization, typically carried out using nature-inspired algorithms. To reduce the CPU expenses, population-based routines are often combined with surrogate modeling techniques, often in the form of machine learning procedures. While offering certain advantages, their efficiency is limited by the curse of dimensionality and antenna response nonlinearity. In this article, we investigate computational advantages of combining population-based optimization with variable-resolution EM models. Toward this end, a model management scheme is developed, which adjusts the discretization level of the antenna under optimization within the continuous spectrum of admissible resolutions. Starting from the lowest usable fidelity, the resolution converges to the highest assumed level when close to the conclusion of the search process. Several adjustment pro-files are considered to investigate the speedup-reliability trade-offs. Numerical results have been obtained for two microstrip antennas and particle swarm optimizer as a representative nature-inspired algorithm. Consistent acceleration of up to eighty percent has been obtained as compared to the single-resolution version with minor degradation of the design quality. Another attractive feature of our methodology is versatility and easy implementation and handling.

ABSTRACT. The measuring system described in the article was designed for measuring large and heavy bent pipes with diameters up to 1.2 m. Currently, measurements of large-diameter bent pipes are taken using either simple and inaccurate protractors or measuring systems requiring a 3D model of the pipe created in graphics software, an optical scanner and markers. This paper presents methods of modeling distortions for measuring system based on one camera that is easy to install in an industrial plant. Those models allow to perform measurement quickly at any position of the pipe on a large industrial measurement table. The paper describes the mathematical models of perspective projections of single- and double- bent pipes, as well as the method of bending angle determination and detection of straight sections of the bent pipe. As part of the research, measurement accuracy of the designed system model were described and confirmed.

ABSTRACT. Thanks to their large number of threads, GPUs allow massive parallelization, hence good performance for numerical simulations, but also make asynchronous execution more common. Kernels that do not actively take part in a computation can be executed asynchronously in the background, in the aim to saturate the GPU threads. We optimized this asynchronous execution by using mixed precision for such kernels. Implemented on the FaSTAR solver and tested on the CRM case, asynchronous execution gave a speedup of 15 to 27 % for a maximum memory overhead of 4.5 to 9 %.

ABSTRACT. Load imbalance is a critical problem that degrades the performance of parallelized applications in massive systems. Although an MPI/OpenMP implementation is widely used for parallelization, users must maintain load balancing at the process level and thread (core) level for effective parallelization. In this paper, we propose dynamic core binding (DCB) to processes for reducing the computation time and energy consumption of applications. Using the DCB approach, an unequal number of cores is bound to each process, and load imbalance among processes is mitigated at the core level. This approach can easy to reduce computation time for applications that have load imbalance. In addition, by reducing the number of cores bound to each process based on the process with the largest load, the energy consumption of applications can be reduced without increasing the computation time. Although load balancing among nodes cannot be handled by DCB, this paper also examines how to solve this problem by mapping processes to nodes. In our numerical evaluations, we implement a DCB library and applied it to applications using lattice H-matrixes, which permit load imbalance in exchange for reduced communication costs. Based on the numerical evaluations, we achieved a 58% performance improvement and 77% energy consumption reduction for the lattice H-matrix using the DCB library.

ABSTRACT. A quantum algorithm for preparing eigenstates of quantum systems is presented which makes use of only forward and backward evolutions under a prescribed Hamiltonian and phase flips. It is based on the Głazek-Wilson-Wegner flow method from condensed-matter physics or more generally double-bracket flows considered in dynamical systems. The phase flips are used to implement a dephasing of off-diagonal interaction terms and evolution reversal is employed for the quantum computer to approximate the group commutator needed for unitary propagation under the double-bracket generator of the diagonalizing flow. The presented algorithm is recursive and involves no qubit overheads. Its efficacy for near-term quantum devices is discussed using numerical examples. In particular, variational double-bracket flow generators, optimized flow step durations and heuristics for pinching via efficient unitary mixing approximations are considered. More broadly, this work opens a pathway for constructing purposeful quantum algorithms based on double-bracket flows also for tasks different from diagonalization and thus enlarges the quantum computing toolkit geared towards practical physics problems.

ABSTRACT. Analysing properties of ground state of a quantum systems, is an im- portant problem with applications in various domains. Recently, Huang et al. [2021] demonstrate how machine learning algorithms can be used to efficiently solve this problem with formal guarantees. However this method requires an ex- ponential amount of data to train. In this work we show a method with improved efficiency for a wide class of energy operator. In particular, we show an ML- based method for predicting ground state properties for structured Hamiltonian with sub-exponential scaling in training data. The method relies on efficiently learning low-degree approximation of the energy operator.

ABSTRACT. A quantum-inspired genetic algorithm can improve a quality of a multi-criteria supervised learning of deep classification models. Designed classifiers can be trained by a quantum simulator with Hadamard gates and rotation gates. To demonstrate advantages of the new algorithm, the algorithm provides the Pareto-optimal classifier for an efficient diagnosis of SARS-CoV-2 virus infection based on remote analysis of X-rays images. Moreover, the quantum processor Sycore-5 was tested with the quantum computing platform QI.

ABSTRACT. As quantum computers mature, the applicability in practice becomes more important. Quantum computers will often be used in a hybrid setting, where classical computers still play an important role in operating and using the quantum computer. However. the term hybrid is diffuse and multi-interpretable. In this work we define two classes of hybrid quantum-classical computing: vertical and horizontal hybrid quantum-classical computing. The first is application-agnostic and concerns using and operating quantum computers. The second is application-specific and concerns running an algorithm. For both, we give a further subdivision in different types of hybrid quantum-classical computing and we introduce terms for them.

ABSTRACT. One of the first fields where quantum computing will likely show its use is optimisation. Many optimisation problems naturally arise in a quadratic manner, such as the quadratic knapsack problem. The current state of quantum computers requires these problems to be for- mulated as a quadratic unconstrained binary optimisation problem, or QUBO. Constrained quadratic binary optimisation can be translated into QUBOs by translating the constraint. However, this translation can be made in several ways, which can have a large impact on the perfor- mance when solving the QUBO. We show six different formulations for the quadratic knapsack problem and compare their performance using simulated annealing. The best performance is obtained by a formulation that uses no auxiliary variables for modelling the inequality constraint.

ABSTRACT. Due to the physics behind quantum computing, quantum circuit designers must adhere to the constraints posed by the limited interaction distance of qubits. Existing circuits need therefore to be modified via the insertion of SWAP gates, which alter the qubit order by interchanging the location of two qubits' quantum states. We consider the Nearest Neighbor Compliance problem on a linear array, where the number of required SWAP gates is to be minimized. We introduce an Integer Linear Programming model of the problem of which the size scales polynomially in the number of qubits and gates. Furthermore, we solve $131$ benchmark instances to optimality using the commercial solver CPLEX. The benchmark instances are substantially larger in comparison to those evaluated with exact methods before. The largest circuits contain up to $18$ qubits or over $100$ quantum gates. This formulation also seems to be suitable for developing heuristic methods since (near) optimal solutions are discovered quickly in the search process.

ABSTRACT. Neural networks perform very well on difficult problems such as image or speech recognition as well as machine text translation. Classification based on fragmented and dispersed data representing certain properties of images or computer's vision is a complex problem. Here, the suitability of a Radial Basis Function (RBF) neural network was evaluated using fragmented data in the problem of recognizing objects in images. The great difficulty of the considered problem is, there isn't an images data as such but only data on some properties of images stored in a dispersed form. More specifically, it was demonstrated that applying a $k-$nearest neighbors classifier in the first step to generate predictions based on fragmented data, and then using a RBF neural network to learn how to correctly recognize the systems of generated predictions for making a final classification is a good approach for recognizing objects in images. In addition, an additional step of training the weights (centers) between the input and hidden layers of a RBF network was proposed. In general, this investigation demonstrates that adding this step significantly improves the correctness of recognizing objects in images.

ABSTRACT. This paper presents a new offline dataset called the Fragments of Medieval Codices (FOMC). It contains medieval Latin handwritings coming from 11th-12th century and can be used to evaluate the performance of offline writer identification and to find the handwriting similarity between the writers or to test the handwritten optical character recognition systems. It consists of 117 fragments of handwritten documents of medieval codices and contains in total over two thousand very high quality images. The collection was assembled using the IIIF standard. We describe the collecting and processing steps performed to develop the dataset and define several evaluation tasks regarding the use of this dataset.

ABSTRACT. Processing of deforming shapes represented by sequences of triangle meshes with connectivity varying in time is difficult, because of the lack of temporal correspondence information, which makes it hard to exploit the temporal coherence. Establishing surface correspondence is not an easy task either, especially since some surface patches may have no corresponding counterpart in some frames, due to self-contact. Previously, it has been shown that establishing sparse correspondence via tracking volume elements might be feasible, however, previous methods suffer from severe drawbacks, which lead to tracking artifacts that compromise the applicability of the results. In this paper, we propose a new, temporally global optimisation step, which allows to improve the intermediate results obtained via forward tracking. Together with an improved formulation of volume element affinity and a robust means of identifying and removing tracking irregularities, the procedure yields a substantially better model of temporal volume correspondence.

ABSTRACT. Deep learning was used to detect boulders that can damage excavators in opencast mines. Different network architectures were applied, e.g., modern YOLOv5. Studies were carried out in which the results obtained using a few networks were compared. In order to further improve the results, a method based on the analysis of the next few frames of the film from an industrial camera was proposed. The results were tested on recordings from a opencast mine in Poland. The proposed method can help prevent the failure of expensive equipment.

ABSTRACT. Human-centric artificial intelligence struggles to build automated procedures that recognize emotions which can be integrated in artificial systems, such as user interfaces or social robots. In this context, this paper researches on building an Emotion Multi-modal Aggregator (EMmA) that will rely on a collection of open-source single source emotion classification programs aggregated to produce an emotion prediction. Although extendable, tested solution takes a video clip and divides into its frames and audio. Then a collection of primary classifiers are applied to each source and their results are combined in a final classifier utilizing machine learning aggregator techniques. The aggregator techniques that have been put to the test were Random Forest and k-Nearest Neighbors which, with an accuracy of 80 \%, have demonstrated superior performance over primary classifiers on the selected dataset.

ABSTRACT. The use of information technologies in industry is growing year by year. More and more advanced devices are implemented and the software needed for them becomes more complex, which increases the risk of errors. To minimize them, it is necessary to constantly monitor the condition of the system and its components. This paper presents a part of a complex production support system for steel mill, responsible for monitoring and tracking the current state on the production hall. Data on currently performed melts and their condition, collected from two sensor layers - Level1 and Level2 - combining with a camera system that allows tracking the position of the main ladle in the hall, was used to create metamodel based on linear regression and neural network for the temperature drop which is occurring during the transport of liquid steel to the casting machine. This approach enables optimization of production volume and minimizes the risk associated with a temperature drop below the optimal one for casting. Several neural network models were used: YOLOv3 for object detection, CRAFT for text detection and CRNN for text recognition. This information is published to the sensor subsystem, enabling precise determination of the state of each performed melt. The system architecture, prediction accuracies and performance analysis were presented.

ABSTRACT. Federated learning is a distributed machine learning method that is well-suited for the Industrial Internet of Things (IIoT) as it enables the training of machine learning models on distributed datasets without the need to centralize the data. One of the most important advantages of using Federated Learning for Automated Guided Vehicles (AGVs) is its capability to optimize resource consumption. AGVs are typically resource-constrained systems and must operate within tight power and computational limits. By using Federated Learning, AGVs can perform model training and updates on-board, which reduces the amount of data that needs to be transmitted. This paper presents experiments on the consumption of resources of the Jetson Nano edge IoT device while training the Federated Learning model and a classical one.

ABSTRACT. We develop a space-time mortar mixed finite element method for parabolic problems modeling flow in porous media. The domain is decomposed into union of subdomains with non-matching grids and different time steps. The space-time variational formulation couples mixed finite elements in space with discontinuous Galerkin in time. Continuity of flux across space-time interfaces is imposed via coarse-scale space-time mortar finite elements, resulting in correct mass conservation. A priori error estimates for the spatial and temporal error are established. A space-time non-overlapping domain decomposition method is developed that reduced the global problem to a space-time coarse-scale mortar interface problem. Each interface iteration requires solving space-time subdomain problems, which is done in parallel. The convergence of the interface iteration is analyzed. Numerical results illustrate the theoretical results and the flexibility of the method for modeling flow in heterogeneous porous media with features localized in space and time.

ABSTRACT. In this talk, we present numerical schemes for both linear and nonlinear poroelasticity problems on convex quadrilateral meshes. We approximate the Darcy pressure using the weak Galerkin (WG) finite element method, and establish discrete weak gradient and numerical velocity in the Arbogast-Correa space. We discuss enriched Lagrangian finite elements and the WG finite element method separately for the elasticity discretization. We formulate the fully-discrete system using implicit Euler time discretization. Iterative methods for dealing the nonlinear cases with dilation- or stress-dependent permeability are discussed. Numerical examples are presented for validating the accuracy and the locking-free property of the new solvers. This is a joint work with Dr. James Liu, Dr. Simon Tavener, and Dr. Ruishu Wang.

ABSTRACT. In this study, we simulated the trajectory of virus-laden droplets from the lung of an infected person to that of the exposed person using computational fluid dynamics. As numerical models, the model of the infected person who had a bifurcated airway and that of the exposed person who had an eighth-generation airway were prepared. The volume and number of virus-laden droplets adhered to the inlet patches of the exposed person's lung were calculated to evaluate the risk of infection when the infected person was talking for 40 seconds. To identify the lung to which droplets adhered, we labeled the inlet patches of the exposed person's lung with 53 numbers, and then measured the volume and number of droplets on the inlet patches of lung. We also categorized the lung's 53 intake patches into five groups and calculated the overall volume and number of attached droplets for each. In addition, we parameterized the angle of the exposed person's neck to evaluate the effect of the tilting neck on the volume and number of droplets reaching the lungs. We found that the volume and number of droplets adhered to the right middle group of bronchi were remarkably smaller than the other four groups, and weakly depended on the neck angle. The volume and number of droplets adhered to the inlet patches of the lung reached the maximum values when the neck angle was 20° upward.

ABSTRACT. Flash calculations of the hydrocarbon mixture are essential for determining how the mixture phase behaves, which will ultimately affect subsurface flow and transport. In this paper, a novel numerical scheme is proposed for calculating the two-phase equilibrium of Peng-Robinson (PR) fluid at constant volume, temperature, and moles, namely the volume-temperature (VT) flash framework based on the dynamic model. Since the dynamic model is based on the energy dissipation law and the Onsager’s reciprocal principle, we proposed a linear energy-stable scheme with the help of the convex-concave splitting technique, the energy factorization approach, and the component-wise iteration framework. The scheme eventually results in a fully explicit algorithm, and it avoids the challenges of solving non-linear systems and other difficulties in the traditional flash calculation methods. This scheme inherits the original energy stability and significantly reduces the implementation burden. It also achieves convergence unconditionally, even with a huge time step. Numerical experiments are carried out to illustrate its accuracy.

ABSTRACT. Ion-induced volume phase transitions in polyelectrolyte gels play an important role in physiological processes such as mucus storage and secretion in the gut, nerve excitation, and DNA packaging. Experiments have shown that changes in ionic composition can trigger rapid swelling and deswelling of these gels. Based on a previously devel- oped computational model, we carry out 2D simulations of gel deswelling within an ionic bath. The dynamics of the volume phase transition are governed by the balance of chemical and mechanical forces on components of the gel. Our simulation results highlight the close connections between the patterns of deswelling, the ionic composition, and the relative magnitude of particle-particle interaction energies.

ABSTRACT. This work presents a Constraint Energy Minimizing Generalized Multiscale Finite Element Method (CEM-GMsFEM) for solving a single-phase compressible flow in highly heterogeneous media. To dis- cretize this problem, we first construct a fine-grid approximation using the Finite Element Method with a backward Euler time approximation. After time discretization, we use Newton’s method to handle the non- linearity in the resulting equations. To solve the linear system efficiently, we shall use the framework of CEM-GMsFEM by constructing multiscale basis functions on a suitable coarse-grid approximation. These basis func- tions are given by solving a class of local energy minimization problems over the eigenspaces that contain local information on heterogeneity. In addition, oversampling techniques provide exponential decay outside the corresponding local oversampling regions. Finally, we will provide two numerical experiments on a 3D case to show the performance of the proposed approach.

ABSTRACT. Smoothed particle hydrodynamics (SPH) is a popular mesh-free Lagrangian method for solving complex fluid flows. In this work, a novel unconditionally energy-stable Smoothed Particle Hydrodynamics (SPH) method is proposed and implemented for incompressible single-phase fluid flow and two-phase fluid flow. For single-phase flow modeled by the Navier-Stokes equation, we apply operator splitting to break the momentum equation into equations involving the non-pressure term and pressure term separately. The idea behind the splitting is to simplify the calculation into a few linear steps while still maintaining unconditional energy stability. With the projection procedure to decouple the momentum and continuity equations, the numerical scheme meets the divergence-free condition. For two-phase flow modeled by the Navier-Stokes–Cahn–Hilliard (NSCH) equation system, we propose a pioneering energy-stable operator splitting strategy to design an efficient Smoothed Particle Hydrodynamics discretization of the Navier-Stokes-Cahn-Hilliard model, which is again unconditionally energy-stable and includes only linear sub-steps.

For both the single-phase flow and two-phase flow, we exploit the great potentials of the Lagrangian particle treatment in the SPH, which have advantages in treating convection-dominated processes. We prove that our SPH method inherits mass and momentum conservation and the energy dissipation properties from the PDE level to the ODE level, and then to the fully discrete level. Consequently and desirably, it also helps increase the stability of the numerical method. Due to its conditional stability, the time step size can be much larger than that of the traditional ISPH methods. This energy-stable SPH method also alleviates the tensile instability without using any particle-shifting strategies, which may destroy the rigorous mathematical proof. Numerical experiments are carried out to show the performance of the proposed energy-stable SPH method for both single-phase and two-phase flows. The inheritance of mass and momentum conservation and the energy dissipation properties are verified numerically. For single-phase flow, numerical examples are presented and compared to the analytical solutions, suggesting that the proposed method has improved accuracy and stability. For two-phase flow, the numerical results also demonstrate that our method captures the interface behavior and the energy variation process well.

ABSTRACT. This study examines the effect of the COVID-19 pandemic and associated lockdowns on access to crucial diagnostic procedures for breast cancer patients, including screenings and treatments. To quantify the impact of the lockdowns on patient outcomes and cost, the study employs a mathematical model of breast cancer progression. The model includes ten different states that represent various stages of health and disease, along with the four different stages of cancer that can be diagnosed or undiagnosed. The study employs a natural history stochastic model to simulate the progression of breast cancer in patients. The model includes transition probabilities between states, estimated using both literature and empirical data. The study utilized a Markov Chain Monte Carlo simulation to model the natural history of each simulated patient over a seven-year period from 2019 to 2025. The simulation was repeated 100 times to estimate the variance in outcome variables. The study found that the COVID-19 pandemic and associated lockdowns caused a significant increase in breast cancer costs, with an average rise of 172.5 million PLN (95% CI [82.4, 262.6]) and an additional 1005 breast cancer deaths (95% CI [426, 1584]) in Poland during the simulated period. While these results are preliminary, they highlight the potential harmful impact of lockdowns on breast cancer treatment outcomes and costs.

ABSTRACT. A growing and renewed interest has emerged in Electronic Health Records (EHRs) as a source of information for decision-making in clinical practice. In this context, the automatic de-identification of EHRs constitutes an essential task, since their dissociation from personal data is a mandatory first step before their distribution. However, the majority of previous studies on this subject have been conducted on English EHRs, due to the limited availability of annotated corpora in other languages, such as Spanish. In this study, we have addressed the automatic de-identification of medical documents in Spanish. A private corpus of 599 real-world clinical cases have been annotated with 8 different protected health information categories. We have tackled the predictive problem as a named entity recognition task, developing two different deep learning-based methodologies, namely a first strategy based on recurrent neural networks (RNN) and an end-to-end approach based on transformers. Additionally, we have developed a data augmentation procedure to increase the number of texts used to train the models. The results obtained show that transformers outperform RNN on the de-identification of Spanish clinical data. In particular, the best performance is obtained by the XLM-RoBERTa large transformer, with a strict-match micro-averaged value of 0.946 for precision, 0.954 for recall and 0.95 for F1-score, when trained on the augmented version of the corpus. The performance achieved by transformers in this study proves the viability of applying these state-of-the-art models in real-world clinical scenarios.

ABSTRACT. Dysgraphia is a learning disability of written expression, which affects the ability to write, primarily handwriting and coherence. Several studies have proposed approaches for assisting dysgraphic people based on AI algorithms. However, existing aids for dysgraphia take only one aspect of the problem faced by the patients into consideration. While some provide writing assis-tance, others address spelling or grammatical problems. In this work, a novel system assisting persons suffering from dysgraphia is proposed. Our system addresses several issues, such as spelling mistakes, grammar mistakes, and poor handwriting. Further, a text-to-speech functionality is added to improve the results. The proposed system binds a plethora of solutions together into a valuable approach for effective writing: a handwritten text recognition using a CNN-RNN-CTC model, a spelling correction model based on SymSpell and Phoneme model, and a grammar correction using the GECToR model. Three machine-learning models are proposed. The result is compared based on the values of Character error rate, Word error rate, and the workflow of three handwritten text recognition models.

ABSTRACT. Balancing Domain Decomposition by Constraints (BDDC) by Dohrmann celebrates 20 years since its publication in 2023. I will present the method and review our research in its extensions to multiple levels, which improves its scalability for large numbers of subdomains and processors. Then, I will present the construction of coarse spaces adapted to the solved problem, which greatly increases the robustness of this iterative method. Together, these form the adaptive-multilevel BDDC, which enjoys both scalability and robustness.Next I will review some applications of the method in computational mechanics. First, I will look at time-dependent Navier-Stokes equations, and I will discus applications of variants of the BDDC method to sequences of Poisson problems arising from the pressure-correction method.  Then, I will describe one path to accelerating the BDDC solver with Graphics Processing Units (GPUs). I will conclude by presenting our recent work towards making multilevel BDDC a robust and scalable method for solving systems arising from immersed boundary finite element methods with adaptive mesh refinement.

ABSTRACT. The paradigm shift towards data-driven science is massively transforming the scientific process. Scientists use exploratory data anal- ysis to arrive at new insights. This requires them to specify complex data analysis workflows, which consist of compositions of data analysis functions. Said functions encapsulate information extraction, integration, and model building through operations specified in linear algebra, rela- tional algebra, and iterative control flow among these. A key challenge in these complex workflows is to understand and act upon irregularities in these workflows, such as outliers in aggregations. Regardless whether irregularities stem from errors or point to new insights, they must be localized and rationalized, in order to ensure the correctness and over- all trustworthiness of the workflow. We propose to automatically reduce a workflow’s input data while still observing some outcome of interest, thereby computing a minimal reproducible example to support workflow debugging. In essence, we reduce the problem to the determination of the input relevant to reproducing the irregularity. To that end, we present a portfolio of different strategies being tailored to data analysis workflows that operate on tabular data. We investigate their feasibility in terms of input reduction, and compare their effectiveness and efficiency within three characteristic cases.

ABSTRACT. A plethora of publicly available, open scholarly data has paved the way for many applications and advanced analytics on science. However, a single dataset often contains incomplete or inconsistent records, significantly hindering its use in real-world scenarios. To address this problem, we propose a framework that allows linking scientific datasets.The resulting connections can increase the credibility of information about a given entity and serve as a link between different scholarly graphs. The outcome of this work will be used in the European Open Science Cloud (EOSC) as a base for introducing new recommendation features.

ABSTRACT. Community detection is a classic problem in network science with extensive applications in various fields. The most commonly used methods are the algorithms designed to maximize modularity over different partitions of the network nodes into communities. Using 80 real and random networks from a wide range of contexts, we investigate the extent to which current heuristic modularity maximization algorithms succeed in returning modularity-maximum (optimal) partitions. We evaluate (1) the ratio of their output modularity to the maximum modularity for each input graph and (2) the maximum similarity between their output partition and any optimal partition of that graph. Our computational experiments involve eight existing heuristic algorithms which we compare against an exact integer programming method that globally maximizes modularity. The average modularity-based heuristic algorithm returns optimal partitions for only 16.9% of the 80 graphs considered. Results on adjusted mutual information show considerable dissimilarity between the sub-optimal partitions and any optimal partitions of the graphs in our experiments. More importantly, our results show that near-optimal partitions tend to be disproportionally dissimilar to any optimal partition. Taken together, our analysis points to a crucial limitation of commonly used modularity-based algorithms for discovering communities: they rarely return an optimal partition or a partition resembling an optimal partition. Given this finding, developing an exact or approximate algorithm for modularity maximization is recommendable for a more methodologically sound usage of modularity in community detection.

ABSTRACT. The task of four-voice harmonization of a given melody is one of the most fundamental, but at the same time the most complex problems in functional harmony. This problem can be formulated as a discrete optimization problem with constraints using a set of rules coming from the theory of music. Unfortunately, a straightforward solution of such a problem, i.e., a mere fulfillment of the rules, ensures only the formal correctness of the obtained chord sequences, which does not necessarily imply overall musical quality as perceived by humans. Trying to catch some non-formalized factors of this quality we have decided to utilize artificial intelligence methods with some ‘creative‘ potential that can provide solutions at acceptable level of formal correctness. In this paper we perform the harmonization using a genetic algorithm, an algorithm based on a Bayesian network, as well as a hybrid of these. In a series of experiments we compare the performance of the three algorithms with each other and with a rule-based system that provides chord sequences at a high level of formal correctness. Besides the formal evaluation all obtained solutions were rated by musical experts. The results show that the studied algorithms can generate solutions musically more interesting than those produced by the rule-based system, even if the former are less formally correct than the latter.

ABSTRACT. The work reported in this article addresses the challenge of building models for non-trivial aerobatic aircraft maneuvers in an automated fashion. It is built using a Behavioural Cloning approach where human pilots provide a set of example maneuvers used by a Machine Learning algorithm to induce a control model for each maneuver. The best examples for each maneuver were selected using a set of objective evaluation metrics. Using those example sets, robust models were induced that could replicate (and in some cases outperform) the human pilots that provided the examples (the clean-up effect). Complete complex maneuvers were performed using a meta-controller capable of sequencing the basic ones learned by imitation. This endeavor was rewarded by the results that show several Machine Learning models capable of performing highly complex aircraft maneuvers.

ABSTRACT. Moral Foundations Theory (MFT) has shown that American liberals and conservatives rely on fundamentally different moral principles, offering a different perspective on the deepening political divide in US politics. However, results outside the US have been less clear, particularly in countries with a more diverse political landscape that does not fall into the traditional Liberal/Conservative dichotomy. Here, we expand the Moral Foundations Dictionary to European Portuguese, which we then use to analyze 10 years of transcripts of parliamentary sessions using standard Data Science and Text Mining techniques. Despite a larger number of represented parties, we show that no traditional parties fall into the Conservative or Liberal characterization and that the political landscape in Portugal is relatively homogeneous with the major difference observed concerning the dichotomy between Government and the parliament.

ABSTRACT. rtificial Intelligence in Education (AIED) offers numerous applications, including student success prediction, which assists educators in identifying the customized support required to improve a student's performance in a course. To make accurate decisions, intelligent algorithms utilized for this task take into account various factors related to student success. Despite their effectiveness, decisions produced by these models can be rendered ineffective by a lack of explainability and trust. Earlier research has endeavored to address these difficulties by employing overarching explainability methods like examining feature significance and dependency analysis. Nevertheless, these approaches fall short of meeting the unique necessities of individual students when it comes to determining the causal effect of distinct features. This paper addresses the aforementioned gap by employing multiple machine learning models on a real-world dataset that includes information on various social media usage purposes and usage times of students, to predict whether they will pass or fail their respective courses. By utilizing Diverse Counterfactual Explanations (DiCE), we conduct a thorough analysis of the model outcomes. Our findings indicate that several social media usage scenarios, if altered, could enable students who would have otherwise received a failing grade to attain a passing grade. Furthermore, we conducted a user study among a group of educators to gather their viewpoints on the use of counterfactuals in explaining the prediction of student success through artificial intelligence.

ABSTRACT. This study aims to analyze sentiment to detect anxiety based on data from social media using machine learning (ML) methods. The dataset was obtained using a crawling process from Indonesian-language YouTube video comments on COVID-19 and the government's program. This topic has many pros and cons that can cause public anxiety and unrest. The data that was successfully collected consisted of datasets and self-collection (4862 data: 3211 negative and 1651 positive). The anxiety is identified in the data based on expert's analysis and grouped into two classes (anxious and not anxious). The identification process uses a machine learning approach with several methods applied, such as KNN (K-Nearest Neighbors), SVM (Support Vector Machine), DT (Decision Tree), Naïve Bayes (NB), RF (Random Forest), and XG-Boost. The data was preprocessed (by tokenizing, filtering, stemming, tagging, and emoticon conversion), and the features were extracted from the results using CV (count-vectorization), TF-IDF (term frequency–inverse document frequency), HV (Hashing-Vectorizer), and Word2Vec (Word Embedding) approaches. The modeling and testing process uses 24 combinations of the ML and feature extraction methods. These experiments aim to get the best performance in anxiety detection, and the combination of RF and CV obtained the best accuracy of 98.4%, which is 13.4 percentage points better than the previous research. In addition, the other ML methods (SVM, DT, RF, and XG-Boost) have accuracy above 92%.

ABSTRACT. Missing data is an issue that can negatively impact any task performed with the available data and it is often found in real-world domains such as healthcare. One of the most common strategies to address this issue is to perform imputation, where the missing values are replaced by estimates. Several approaches based on statistics and machine learning techniques have been proposed for this purpose, including deep learning architectures such as generative adversarial networks and autoencoders. In this work, we propose a novel siamese neural network suitable for missing data imputation, which we call Siamese Autoencoder-based Approach for Imputation (SAEI). Besides having a deep autoencoder architecture, SAEI also has a custom loss function and triplet mining strategy that are tailored for the missing data issue. The proposed SAEI approach is compared to seven state-of-the-art imputation methods in an experimental setup that comprises 14 heterogeneous datasets of the healthcare domain injected with Missing Not At Random values at a rate between 10% and 60%. The results show that SAEI significantly outperforms all the remaining imputation methods for all experimented settings, achieving an average improvement of 35%.

ABSTRACT. Data from Social Network is a valuable asset within both a scientific and a business world. In the context of this work, ontological modelling from Social Network is understood as a knowledge building process to generate a shared domain model. Such a technique relies on a balanced co-existence of human intuition/creativity and technological support. We assume collaborative modelling and collective/social intelligence. It implies a certain degree of uncertainty that is, in principle, inversely proportional to the achieved consensus. There are two clear different convergence points between the proposed process and collective/social intelligence: (i) at a data level, because of the nature of the input which is generated by different individuals, communities, stakeholders and actors; and (ii) at a modelling level, where human inputs, design decisions and validations are expected to involve several contributors, experts, modellers or analysts. Although looking holistically at the modelling process, this paper concisely focuses on the Ontology and the associated uncertainty, while resulting systems and studies are object of future work.

ABSTRACT. In this paper, the specific kind of uncertainties, which appear in ontology-driven software development, is discussed. We focus on the development of IoT applications whose source code is generated automatically by means of some framework as an ontology-driven solution. So-called "compatibility uncertainties" pop up when the ontology has been changed while the corresponding generated application is in operation. This specific kind of uncertainties can be treated as a variant of implementation uncertainties. The algorithm of its automated handling is presented. The proposed algorithm is implemented within the SciVi platform and tested in the real-world project devoted to custom IoT-based hardware user interfaces development for virtual reality. We use the SciVi platform as a toolset for the automatic generation of IoT devices firmware for ontology-driven Edge Computing but the problem discussed is common for any tools which are used for the generation of ontology-driven software.

ABSTRACT. Variance- and density-based approaches are widely used for global sensitivity analysis (GSA) of system dynamics and agent-based models (ABM). However, these techniques are limited in cases where a comprehensive understanding of temporal dynamics is critical. This is particularly relevant to models that exhibit diverse dynamics on varying timescales and with varying degrees of structural complexity. To address this, we propose a novel manifold learning-based method for GSA in systems exhibiting complex spatiotemporal processes. Our method involves identifying the embedding of high-dimensional output data of a computational model using Grassmannian diffusion maps, which allows us to reduce the dimensionality of the data and identify meaningful geometric descriptions in a parsimonious manner. We then use polynomial chaos expansion (PCE) to construct a mapping between the stochastic input parameters and the diffusion coordinates of the reduced space. Finally, we calculate sensitivity indices using the PCE coefficients. We demonstrate the technique's capabilities by applying it to the traditional Lotka-Volterra model and an ABM of epidemic dynamics. Unlike traditional GSA methods that require evaluations at multiple time steps and for several model outputs, our method aggregates entire trajectories of multiple model responses, providing a more general estimation of parameter sensitivities. Additionally, our approach captures slow and fast processes in the model, making it a useful tool for systems with diverse temporal dynamics. Furthermore, we establish that our proposed method meets all the "good" properties of a global sensitivity measure, making it a valuable alternative to variance- and density-based GSA. We anticipate that our technique will expand the use of manifold-based approaches and contribute to a deeper understanding of complex spatiotemporal processes.

ABSTRACT. Determining what causes field equipment malfunction and predicting when those malfunctions will occur can save large amounts of money to corporations that are capital intensive. To prevent equipment down time, maintenance field equipment maintenance departments must be correctly dimensioned. Herein, we demonstrate the efficacy of machine learning to determine time between failure, repair time (equipment down time) and repair cost. Also, a mean value analysis is carried out to determine the maintenance department capacity. Uncertainty is modelled using statistical analysis and simulation.

ABSTRACT. In this paper, we propose an approach for an application of Bayesian optimization using Sequential Monte Carlo (SMC) and concepts from the statistical physics of classical systems. Our method leverages the power of modern machine learning libraries such as NumPyro and JAX, allowing us to perform Bayesian optimization on multiple platforms, including CPUs, GPUs, TPUs, and in parallel. Our approach enables a low entry level for exploration of the methods while maintaining high performance. We present a promising direction for developing more efficient and effective techniques for a wide range of optimization problems in diverse fields.

ABSTRACT. The problem of fitting a given ordered sample of data points in arbitrary Euclidean space is addressed. The corresponding interpolation knots remain unknown and such must be first somehow found. The latter leads to a highly non-linear multivariate optimization task, equally non-trivial for theoretical analysis and derivation of a computationally efficient numerical scheme. The non-degenerate case of at least four data points can be handled by Leap-Frog algorithm merging generic and nongeneric univariate overlapping optimizations. Sufficient conditions guaranteeing the unimodality for both Leap-Frog optimization tasks are already established in the previous research. This works complements the latter by analyzing the degenerate case i.e. the case of fitting three data points, for which Leap-Frog cannot be used. It is proved here that the related univariate cost function is always unimodal yielding a global minimum assigned to the missing middle-point knot (both terminal knots can be assumed to be fixed). Illustrative examples supplement the analysis in question.

ABSTRACT. This work presents a research on Nature Inspired Metaheuristic Algorithms (MA) used as optimizers in training process of Machine Learning method called Extreme Learning Machine (ELM). We tested 19 MA optimizers measuring their performance directly on sample datasets. The impact of input parameters such as number of hidden layer units, optimization stopping conditions and population size on the accuracy results, training and prediction time is evaluated here. Significant differences in performance of applied methods and their parameters' values are detected. The most meaningful outcome of this paper shows that an increase of the number of MA iterations does not yield significant boost in accuracy with a huge increase in training time. Indeed a limit from 1 to 5 iterations of MA is sufficient for analyzed machine learning tasks. In our research the best results are obtained for population size ranging between 50 and 100. Hybridized ELM outperforms classical implementation of ELM for which higher accuracy is reached for the same number of neurons.

ABSTRACT. We present an improved efficiency Leading Zero Counter for Xilinx FPGAs which improves the path delay while maintaining the re- source usage, along with generalizing the scheme to variants whose in- puts are of any size. We also show how the Ultrascale architecture also allows for better Intellectual Property solutions of certain forms of this circuit with its newly introduced logic elements. We also present a de- tailed framework that could be the basis for a methodology to measure results of small-scale circuit designs synthesized via high-level synthesis tools. Our result shows that very high frequencies are achievable with our design, especially at sizes where common applications like floating point addition would require them. For 16, 32 and 64-bit, our real-world build results show a 6% 14% and 19% path delay improvement respectively, enough of an improvement for large scale designs to have the possibility to operate close to the maximum FPGA supported frequency.

ABSTRACT. Regression of potential energy functions is one of the most popular applications of machine learning within the field of material simulations since it would allow accelerating molecular dynamics simulations. Recently, graph-based architectures have been proven to be especially suitable for molecular systems. However, the construction of robust and transferable potentials, resulting in stable dynamical trajectories, still needs to be researched. In this work, we design and compare several neural architectures with different graph embedding layers to predict the energy of water cluster anions, a system of fundamental interest in chemistry and biology. After identifying the best aggregation procedures for this problem, we have obtained accurate, fast-evaluated and easy-to-implement graph neural network models which could be employed in dynamical simulations in the future

ABSTRACT. Abrupt and rapid high-amplitude changes in a dynamical system’s states known as extreme events appear in many processes occurring in nature, such as drastic climate patterns, rogue waves, or avalanches. These events often entail catastrophic effects, therefore their description and prediction is of great importance. However, because of their chaotic nature, their modelling represents a great challenge up to this day. The applicability of a data-driven modularity-based clustering technique to identify precursors of rare and extreme events in chaotic systems is here explored. The proposed identification framework based on clustering of system states, probability transition matrices and state space tessellation was developed and tested on two different chaotic systems that exhibit extreme events: the Moehliss-Faisst-Eckhardt model of self-sustained turbulence and the 2D Kolmogorov flow. Both exhibit extreme events in the form of bursts in kinetic energy and dissipation. It is shown that the proposed framework provides a way to identify pathways towards extreme events and predict their occurrence from a probabilistic standpoint. The clustering algorithm correctly identifies the precursor states leading to extreme events and allows for a statistical description of the system’s states and its precursors to extreme events.

ABSTRACT. This paper proposes a wildfire prediction model, using machine learning, social media and geophysical data sources to predict wildfire instances and characteristics with high accuracy. We use social media as a predictor of wildfire ignition, and a machine learning based reduced order model as a fire spead predictor. We incorporate social media data into wildfire instance prediction and modelling, as well as leveraging reduced order modelling methods to accelerate wildfire prediction and subsequent disaster response effectiveness.

ABSTRACT. The study concerns classification based on dispersed data, more specifically data collected independently in many local decision tables. The paper proposes an approach in which coalitions of local tables are generated using Pawlak's conflict analysis model. Decision trees are built based on tables aggregated within the coalitions The paper examines the impact of the stop criterion (determined by the number of objects in a node) on the quality classification and on the rules' quality generated by the model. The results are compared with the baseline approach, in which decision trees are built independently based on each of the decision tables. The paper shows that using the proposed model, the generated decision rules have much greater confidence than the rules generated by the baseline method. In addition, the proposed model gives better classification quality regardless of the stop criterion compared to the non-coalitions approach. Moreover, the use of higher values of the stop criterion for the proposed model significantly reduces the length of the rules while maintaining the classification accuracy and the rules' confidence at a high level.

ABSTRACT. Turbulence is characterised by chaotic dynamics and a high-dimensional state space, which make the phenomenon challenging to predict. However, turbulent flows are often characterised by coherent spatiotemporal structures, such as vortices or large-scale modes, which can help obtain a latent description of turbulent flows. However, current approaches are often limited by either the need to use some form of thresholding on quantities defining the isosurfaces to which the flow structures are associated or the linearity of traditional modal flow decomposition approaches, such as those based on proper orthogonal decomposition. This problem is exacerbated in flows that exhibit extreme events, which are rare and sudden changes in a turbulent state. The goal of this paper is to obtain an efficient and accurate reduced-order latent representation of a turbulent flow that exhibits extreme events. Specifically, we employ a three-dimensional multiscale convolutional autoencoder (CAE) to obtain such latent representation. We apply it to a three-dimensional turbulent flow. %in which quasi-relaminarization events with their bursts in kinetic energy and dissipation rate are the extreme events. We show that the Multiscale CAE is efficient, requiring less than 10% degrees of freedom than proper orthogonal decomposition for compressing the data and is able to accurately reconstruct flow states related to extreme events. The proposed deep learning architecture opens opportunities for nonlinear reduced-order modeling of turbulent flows from data.

ABSTRACT. Systems with time-delayed chaotic dynamics are common in nature, from control theory to aeronautical propulsion. The overarching objective of this paper is to compute the stability properties of a chaotic dynamical system, which is time-delayed. The stability analysis is based only on data. We employ the echo state network (ESN), a type of recurrent neural network, and train it on timeseries of a prototypical time-delayed nonlinear thermoacoustic system. By running the trained ESN autonomously, we show that it can reproduce (i) the long-term statistics of the thermoacoustic system's variables, (ii) the physical portion of the Lyapunov spectrum, and (iii) the statistics of the finite-time Lyapunov exponents. This work opens up the possibility to infer stability properties of time-delayed systems from experimental observations.

ABSTRACT. Many machine learning models are applied on facial expression classification and there are three main issues affecting the performance of any algorithms in classifying emotions based on facial expressions, and these issues include image illumination, image quality and partial features recognition. Many approaches have been proposed to handle these issues. Unfortunately, one of the main challenges in detecting and classifying facial expression process is minimal differences of features between different types of emotions that can be used to differentiate these different types of emotions. Thus, there is a need to enrich each type of emotion with more relevant extracted features by having a more effective approach to extract features that can be used to represent each type of emotions more effectively and efficiently. This work addresses the issue of improving the emotion recognition accuracy by introducing a novel hybrid approach that combines the Depth Active Appearance Model (DAAM) and Deep Convolutional Neural Networks (DCNNs). The proposed DAAM and DCNNs model can be used to assist one in identifying emotions and classify learner involvement and interest in the topic which are plotted as feedback to the instructor to improve learner experience. The proposed method is evaluated on two publicly available datasets namely, JAFFE and CK+ and the results are compared to the state-of-the-art results. The empirical study showed that the proposed DAMM-CNNs hybrid method managed to perform the face expression recognition with 97.4\% for the JAFFE dataset and 96.9\% for the CK+ dataset.

ABSTRACT. Extreme Learning Machine (ELM) is a feed-forward neural network with one hidden layer. In its modification called ELM Radial Basis Function the input data is a priori clustered into a number of sets represented by their centroids. The matrix of distances between each sample and centroid is calculated and applied as input data to the neural network. This work conducts a comparison study of the ELM Radial Basis Function classification performance upon applying either k-means, k-medoids or mean shift clustering methods. Generated results are obtained from two datasets i.e. Wine Quality-White and Ionosphere. The computations are based on full datasets or on the same sets reduced by a feature selection algorithm. The parameters of the classifiers such as number of neurons in hidden layer, value of k in k-means and k-medoids, value of radius in mean shift are optimized through an iterative procedure maximizing an accuracy or minimizing Mean Square Error and computation time. The different distance metrics for k-means and k-medoids, and mean shift with gaussian or flat kernel function are also compared. The results obtained with Softplus and linear activation function (applied to the net in most of the computations in this work) are juxtaposed with the results generated by other activation functions.

ABSTRACT. Natural language processing, like other ML, DL, and data processing tasks, requires a large amount of data to be effective. Thus, one of the most significant challenges confronting ML/DL tasks, including NLP, is a lack of data. This is especially noticeable in the case of text data for niche languages like Polish. Manual collection and labelling of text data is the primary method for obtaining language-specific data. However, this is a lengthy and labour-intensive process. As a result, researchers use a variety of other solutions, such as machine translation from another, more developed language in the field, to obtain data more quickly and affordably. Most commonly, this is English, as it is the universal language used by researchers worldwide. Experiments were carried out, resulting in a machine translation model that translates texts from English into Polish. Its results were compared with those of a pre-trained model and translations were subjected to human testing. The influence of different pre-processing stages on the final result of the classification of texts in Polish in terms of one of the six emotions contained in them was also checked: anger, fear, joy, love, sadness, surprise.

ABSTRACT. This paper presents a new method for estimating the radius of blood vessels using vesselness functions computed at multiple scales. The multiscale vesselness technique is commonly used to enhance blood vessels and reduce noise in angiographic images. The corrected and binarized image resulting from this technique is then used to construct a 3D vector model of the blood vessel tree. However, the accuracy of the model and consequently the accuracy of radii estimated from the model may be limited by the image voxel spacing. To improve the accuracy of the estimated vessel radii, the method proposed in this study makes use of the vesselness functions that are already availabe as by-products of the preceding enhancement procedure. This approach speeds up the estimation process and maintains sub-voxel accuracy. The proposed method was validated and compared with two other state-of-the-art methods. The quantitative comparison involved artificially generated images of tubes with known geometries, while the qualitative assessment involved analyzing a real magnetic resonance angiogram. The results obtained demonstrate the high accuracy and usefulness of the proposed method. The presented algorithm was implemented, and the source code was made freely available to support further research.

ABSTRACT. With the popularity of web applications, web attacks have become one of the major threats to cyberspace security. Many studies have focused on applying machine learning techniques for web attack traffic detection. However, past approaches suffer from two shortcomings: firstly, handcrafted feature extraction-based approaches cannot adapt to changing attack patterns, and secondly, current end-to-end deep learning approaches treat traffic payloads as non-structured string sequences ignoring their inherent structural characteristics. Therefore, we propose a graph-based web attack traffic detection model to identify the payloads in the traffic requests. Each pre-processed payload is transformed as an independent graph in which the node representations are shared through a global feature matrix. Finally, graph-level classification models are trained with graph attention networks combining global information. Experimental results on four publicly available datasets show that our approach successfully exploits local structural characteristics and global information to achieve state-of-the-art performance.

ABSTRACT. Obstacles pose serious risks and dangers for blind and visually impaired individuals (BVI) especially when no companion or assistant is present. Hence, we propose a head-mounted edge computing device prototype is proposed in this study to address this challenge. The aim is to establish a computationally efficient mechanism that can accurately warn the risk if there is an obstacle on the path. The learned obstacle warning model needs to be reliable yet small in size so it can be embedded in the wearable device and run without consuming much energy. In addition, it needs to deal with natural head turns which have a serious impact on readings from head-mount sensors. More than thirty models have been investigated in order to determine the most appropriate model which can balance the accuracy and real-time performance by comparing the key metrics of these models. This study shows that a highly efficient wearable device is feasible and can help BVI people avoid obstacles with high accuracy. In addition, a large data set has been collected which can be used as a benchmark for related studies.

ABSTRACT. As network scale expands and concurrent requests grow, unexpected network anomalies are more frequent, leading to service interruptions and degraded user experience. Real-time, accurate troubleshooting is critical for ensuring satisfactory service. Existing troubleshooting solutions adopt ensemble anomaly detection (EAD) to detect anomalies due to its robustness. However, the fixed base classifier parameters in EAD set by expert experience may reduce the efficiency of anomaly detection when faced with different data distributions. Furthermore, the binary results fed to the secondary classifier in EAD cause information loss, leading to compromised accuracy and inaccurate root cause localization. Additionally, the use of multiple redundant KPI data to identify the root causes of anomalies is time-consuming and error-prone. To address the above issues, we propose a fully automatic troubleshooting system, ATS. The new EAD method AutoDetect is introduced to detect anomalies, then Gunlock is designed to trigger AutoRoot to locate the root cause. Specifically, AutoDetect updates the parameters of base classifiers to dynamically adapt to different KPI data distributions. The ensemble of soft labels generated by base classifiers is subsequently fed into the secondary classifier to achieve information-lossless anomaly detection. Finally, a heuristic module Gunlock is proposed to select the most appropriate KPI data based on the metric i.e., bilayer relative difference to trigger AutoRoot for efficient root cause localization. Extensive experiments on multiple real traces demonstrate that ATS is more than twice as fast as most state-of-the-art solutions while with higher troubleshooting accuracy.

ABSTRACT. The performance and range of electric vehicles are largely determined by the capabilities of their battery systems. To ensure optimal operation and protection of these systems, Battery Management Systems rely on key information such as State of Charge, State of Health, and sensor readings. These critical factors directly impact the range of electric vehicles and are essential for ensuring safe and efficient operation over the long term. This paper presents the development of a battery State of Charge estimation model based on a 1-D convolutional neural network. The data used to train this model are theoretical operating data as well as driving cycles of lithium batteries. An Explainable Artificial Intelligence method is then applied to this model to verify the physical behavior of the black box model. Finally, a testing platform was constructed to assess the effectiveness of the State of Charge estimation model. Our explainable model, called SocHAP, is compared against other contemporary methods to evaluate its predictive accuracy.

ABSTRACT. Poly lactic-co-glycolic acid (PLGA) is a copolymer that have demonstrated great potentials in development of novel drug delivery systems. This paper first discusses synthesis procedures and properties of PLGA nanoparticles (NPs) and then examine mechanisms of drug release from PLGA nanoparticles. For the core-shell structure of reservoir-type PLGA NPs, diffusion through the polymeric shell is identified as the main mechanism of drug release. A time-dependent diffusion equation is used to model release from homogeneous spherical particles. A mass-conservative numerical method for the radial diffusion model is developed for numerical simulations. Comparison with in vitro results demonstrates usefulness of numerical simulations.

ABSTRACT. The paper presents the results of numerical simulation of the propeller Ka4-70 using the actuator line model in the OpenFOAM, AMReX and Nek5000 open-source CFD softwares. The modification of the tools for wind farm simulation for these packages has been carried out. Features of these implementations are described. For numerical calculations the IDDES turbulence model is used. A comparison of the computational costs and accuracy of flow structures were made for the actuator line model using different open-source CFD softwares and the arbitrary mesh interface method. The actuator line model provides force characteristics and flow structures with good enough accuracy.

ABSTRACT. It is preferrable to use quadrilateral and hexahedral grids for the computation of fluid flow and transport in porous media typically arising from geological applications. The standard mixed finite element method (FEM) usually leads to a saddle-point system, which may not be suitable for large-scale computing. An alternative method is the least-squares finite element method (LSFEM). It satisfies the coercivity. This is a big advantage. Among others, the subtle issue on how to choose the finite element space in the other FEMs is now avoided, i.e., any finite element space can be used. On the other hand, LSFEMs still encounter some common problems as in the standard mixed FEM. The problems we are considering are that the nonaffine grids such as quadrilaterals and hexahedra are usually not valid. This is often the case for the Raviart-Thomas quadrilateral elements and Raviart-Thomas-Nedelec hexahedral elements. The main reason lies in that the highly distorted meshes resulting from the so-called Piola transformation from the reference square or cubic lead to suboptimally and even non convergent approximations. In numerical analysis, this is due to the fact that there lacks the so-called commuting diagram property among the associated finite element spaces. We here will present a new LSFEM. The new feature is that we apply some elemental local L2 projectors to some terms of the residuals from the original partial differential equations. Then, with the discovery that the Raviart-Thomas quadrilateral elements and Raviart-Thomas-Nedelec hexhedral elements satisfy a very important property: as we called, the quasi-commuting diagram property among the local finite element spaces of polynomials over the reference element (square or cubic). The local L2 projections make full use of such property so that all Raviart-Thomas quadrilateral elements are optimally convergent and so that except for the lowest-order element, all the Raviart-Thomas-Nedelec hexahedral elements can be optimally convergent. We must point out that compared with other H(div) elements on affine grids and with other FEMs such as the so-called multipoint flux mixed FEMs, the advantages of the original Raviart-Thomas quadrilateral elements and Raviart-Thomas-Nedelec hexahedral elements and our new LSFEM have several important aspects. One is that as far as the hexahedral grids are concerned, till today, Raviart-Thomas-Nedelec hexahedral elements are still the simplest elements through the Piola transformation from the reference element. Second is that our new LSFEM is optimally convergent for all general quadrilaterals and hexahedra, except for the lowest-order hexahedral element. The hexahedra are allowed to have general faces such as curved faces. Only assume the standard shape-regularity on the meshes. Not any additional assumptions are made. Three is that our new LSFEM is still coercive, resulting in a symmetric positive definite system, even if the original problem is non symmetric and indefinite. Four is that our new LSFEM is suitable for convection-dominated problems. To the authors’ knowledge, few is known in literature for other LSFEMs which can effectively and efficiently solve the convection-dominated problems. Five is that our new LSFEM satisfies a so-called norm equivalence with respect to the finite element spaces so that a simple and natural diagonal block preconditioner is provided, and consequently, with such preconditioner, the application of the preconditioned Conjugate Gradient algorithm yields a convergence uniform in the mesh size. All these would find very important merits in large-scale computations in practice, e.g., the applications to the numerical modeling of carbon sequestration in deep saline aquifers and to the numerical modeling complex multiphase multicomponent fluid flows in porous media

ABSTRACT. Direct Numerical Simulation of thermocapillary-migration of a bi-dispersed suspension of droplets is performed using multiple markers unstructured conservative level-set method for two-phase flow with variable surface tension. Surface tension is a function of temperature on the interface. Consequently, the called Marangoni stresses induced by temperature gradients on the interface lead to a coupling of the momentum transport equation with the thermal energy transport equation. The finite-volume method on 3D collocated unstructured meshes is used to discretize the transport equations. The unstructured conservative level-set method is employed for interface capturing, whereas the multiple marker approach avoids the numerical coalescence of droplets. The pressure-velocity coupling is solved by the fractional-step projection method. Unstructured flux limiters approximate the convective term of transport equations. Adaptive mesh refinement is introduced for the optimization of computational resources. Verifications, validations and numerical findings are reported.

ABSTRACT. The paper is devoted to the study of various numerical algorithms for calculating the flow and acoustics characteristics of supersonic jets implemented in open source softwares. The ideally expanded supersonic jet with parameters M = 2.1, Re = 70000 is considered. A comparison of various approaches implemented in the OpenFOAM and block-structured adaptive mesh refinement framework of AMReX is made. Numerical algorithms for compressible gas flow implemented in pimpleCentralFoam, QGDFoam and CNS solvers are considered. Acoustic noise are calculated using the Ffowcs Williams and Hawkings analogy implemented in the libAcoustics library. Cross-validation comparison of the flow fields and acoustic characteristics has been carried out.

ABSTRACT. This study presents a comprehensive comparative analysis of several meta-heuristic algorithms for solving the damage detection problem in structures. The problem is modelled as a bounded single objective optimization problem, and the performance and efficiency of the algorithms are compared under various scenarios using noise-contaminated data. The study results show that these meta-heuristics are powerful methods for global optimization and are suitable for solving the damage detection problem. The study also compares the performance of these algorithms in terms of the ability to identify the location and severity of the damage, as well as the robustness of the algorithms to noise-contaminated data. The study concludes that the proposed meta-heuristic algorithms are promising techniques for damage detection, and show that the GSK_ALI, MRFO, and Jaya algorithms demonstrate superior performance when compared to the other algorithms in the evaluation to detect damaged elements in the structures.

ABSTRACT. Rare autoimmune diseases provoke immune system malfunctioning, which reacts and damages the body's cells and tissues. They have a low prevalence, classified as complex and multifactorial, with a difficult diagnosis. In this sense, this work aims to support the diagnosis of a rare autoimmune disease using the analysis of medical records from supervised machine learning methods and to identify the models with the best performance considering the characteristics of the available data set. A synthetic database was created with 1000 samples from epidemiologi-cal studies identified in the literature, simulating demographic data and symptoms from patient records for the diagnosis of Amyotrophic Lateral Sclerosis (ALS), Multiple Sclerosis (MS), Systemic Lupus Erythematosus (SLE), Crohn's Disease (CD) and Autoimmune Hepatitis (AIH). Data were segmented into training (80%) and test (20%), assigning the diagnosis as the class to be predicted. The models with the highest accuracy were Logistic Regression (84%), Naive-Bayes (82.5%), and Support Vector Machine (79%). Only RL obtained quality values greater than 0.8 in all metrics. LES had the highest precision and recall for all classifiers, while ELA had the worst results. The RL model had the best perfor-mance with good quality metrics, although it was impossible to predict ALS ac-curately. Although a real dataset was not used, this work presents a promising approach to support the challenging diagnosis of rare autoimmune diseases.

ABSTRACT. n this article, we present a processing pattern dedicated to video-oculography. It is a complete solution that allows for checking the external conditions accompanying the video oculography, conducting oculometric measurements based on different test models, estimating eye movement (EM) angles and detecting EM type in the stream of coordinates, and calculating its parameters. We based our architecture on neural networks (NN), machine-learning (ML) algorithms of different types, parallel/asynchronous computing, and compilations of the models, to achieve real-time processing in oculometric tests.

Oculometric tests provide significant insight into central neuro-motor states, but machine-learning methods are needed to estimate their meaning. A limitation of this cognitive-analytical trend was the reliance on dedicated measuring devices, such as eye-trackers, which are usually separate and expensive equipment. Our approach goes beyond these limitations and was developed to use standard home computer equipment, with an internet connection and computer camera. Our set of dedicated algorithms embedded in the software compensates for hardware limitations.

We tested the results of the presented solution on reflexive saccades (RS) and a standard 30 frames per second (FPS) web-camera and were able to distinguish between young and old persons and between healthy and prodromal ND subjects by analyzing RS parameters. Visual processes are connected to many brain structures and, by using ML methods, we are trying to dissect them. The development of solutions like the one presented in this article gives hope for the general availability of screening tests connected to ND risk and for collecting extensive data outside the laboratory.

We hope that this direction will contribute to the development of ND analytic means in computational health and consequently, to the faster development of new ND preventive measures.

ABSTRACT. Diabetic Foot Ulcers (DFUs) are a severe complication of Diabetes Mellitus. The development of a DFU can cause a sharp decline in a patient's health and quality of life, often leading to further infection, amputation and death. The process of risk stratification is therefore crucial for informing the level and regularity of care that a patient should receive to help manage their Diabetes before a DFU can form. In existing practice, risk stratification is often a manual process where a clinician allocates a risk category based on a patient’s biomarker features captured during routine appointments. In this paper, we present the initial outcomes of a feasibility study on the use of data analysis and machine learning techniques for the purpose of risk stratification of DFU formation using biomarker data. We explore this through application of Machine Learning (ML) algorithms to patient data collected during routine medical appointments.

ABSTRACT. The incredible advances in machine learning have created a variety of predictive and decision-making medical models that greatly improve the efficacy of treatment and improve the quality of care. In healthcare, such models are often based on electronic health records (EHRs). The quality of this models depends on the quality of the EHRs, which are usually presented as plain unstructured text. Such records often contain spelling errors, which reduce the quality of intelligent systems based on them. In this paper we present a method and tool for correcting spelling errors in medical texts in Russian. By combining the Symmetrical Deletion algorithm and a finely tuned BERT model to correct spelling errors, the tool can improve the quality of original medical texts without significant cost. We have evaluated the correction precision and performance of the presented tool and compared it with other popular spelling error correction tools that support Russian language. Experiments have shown that the presented approach and tool are outperform to existing open source tools for automatic spelling error correction in Russian medical texts. The proposed tool and its source code are available on GitHub (https://github.com/DmitryPogrebnoy/MedSpellChecker) and pip (https://pypi.org/project/medspellchecker) repositories.

ABSTRACT. Obtaining overlaps between all pairs from billions of strings is a fundamental computational challenge in de novo whole genome assembly. This paper presents a memory-efficient, massively parallel algorithm that uses GPUs to find overlaps from large sequencing datasets with a very low probability of false positives. Here we use a Rabin-fingerprint-based indexing method which stores the strings with their fingerprints and uses them to generate the fingerprints of suffixes and prefixes. We then sort these fingerprints in hybrid CPU-GPU memory and stream them in GPU to find matches. This approach is also amenable to multi-GPU and out-of-core processing. Experiments show that our implementation can detect a trillion highly probable overlaps from within 1.5 billion DNA fragments in just over two hours. Compared to the existing CPU-based approach, it routinely achieves speedups of 5-15x while having an error rate of 1 in 20 million.

ABSTRACT. The adjoint reversal of long evolutionary calculations (e.g. loops), where each iteration depends on the output of the previous iterate, is a common occurrence in computational engineering (e.g. computational fluid dynamics (CFD) simulation), physics (e.g. molecular dynamics) and computational finance (e.g. long Monte Carlo paths). For the edge case of a scalar state, the execution, as well as adjoint control flow reversal, are inherently serial operations, as there is no spatial dimension to parallelize. Our proposed method exploits the run time difference between passive function evaluation and augmented forward evaluation, which is inherent to most adjoint AD techniques. For high dimensional states, additional parallelization of the primal computation can and should be exploited at the spatial level. Still, for problem sizes where the parallelization of the primal has reached the barrier of scalability, the proposed method can be used to better utilize available computing resources and improve the efficiency of adjoint reversal.

We expect this method to be especially useful for operator overloading AD tools. However, the concepts are also applicable to source-to-source transformation and handwritten adjoints, or a hybrid of all approaches. For illustration, a C++ reference implementation of a scalar evolution reversal is presented utilizing handwritten adjoint code.

ABSTRACT. This paper presents the results of a preliminary experimental investigation of the performance of a stationary iterative method based on a block staircase splitting for solving singular systems of linear equations arising in Markov chain modelling. From the experiments presented, we can deduce that the method is well suited for solving block banded or more generally localized systems in a parallel computing environment. The parallel implementation has been benchmarked using several Markovian models.

ABSTRACT. Modeling large datasets through Multivariate Functional Approximations (MFA) plays a critical role in scientific analysis and visualization workflows. However, this requires scalable data partitioning approaches to compute MFA representations in a reasonable amount of time. We propose a fully parallel and efficient method for computing MFA with B-spline bases without sacrificing the reconstructed solution accuracy or continuity. Our approach reduces the total work per task and uses a restricted Additive Schwarz (RAS) method to converge control point data across subdomain boundaries. We also provide a detailed analysis of the parallel approach with domain decomposition solvers to minimize subdomain error residuals and recover high-order continuity with optimal communication cost determined by the overlap regions in the RAS implementation. In contrast to previous methods that generally only recover $C^1$ continuity for arbitrary B-spline order $p$ or required post-processing to blend discontinuities in the reconstructed data, the accuracy of the MFA remains bounded as the number of subdomains is increased. We demonstrate the effectiveness of our approach using analytical and scientific datasets in 1, 2, and 3 dimensions and show that it is highly scalable (due to bounded outer iteration counts) and that the parallel performance at scale is directly proportional to the nearest-neighbor communication implementations.

ABSTRACT. Tempo and time signature detection are essential tasks in the field of Music Information Retrieval. These features often affect the perception of a piece of music. Their automatic estimation unlocks many possibilities for further audio processing, as well as supporting music recommendation systems and automatic song tagging. In this article, the main focus is on building a two-phase system for extracting both features. The influence of many parameters of known methods was investigated. The results were also compared with the well-known and scientifically recognized Librosa library for music processing.

ABSTRACT. This article introduces an easy to implement kick-starting method for transfer learning of image recognition models, meant specifically for training with limited computational resources. The method has two components: (1) Principal Component Analysis transformations of per-filter representations and (2) explicit storage of compressed features. Apart from these two operations, the latent representation of an image is priorly obtained by transforming it via initial layers of the base (donor) model. Taking these measures saves a lot of computations, hence meaningfully speeding up the development. During further work with models, one can directly use the heavily compressed features instead of the original images each time. Despite having a large portion of the donor model frozen, this method yields satisfactory results in terms of prediction accuracy. Such a procedure can be useful for speeding up the early development stages of new models or lowering the potential cost of deployment.

ABSTRACT. Biometric identification from heart rate sequences provide a simple yet effective mechanism, that can neither be reverse engineered nor replicated, to protect user privacy. This study employs a highly efficient time series classification (TSC) algorithm, miniROCKET, to identify users by their heart rate. The approach adopted in this study employs user heart rate data, a simplified form of heart activity, captured during exercise, filtered, and contextualized within exercise routines, for user classification. Results from this study are empirically evaluated against three other state-of-the-art TSC algorithms, on a real-world data set containing 115,082 workouts across 304 users of IoT wearables. Our experiments showed that for 36 users, the variance explained by the heart rate feature alone was 74%, when coupled with speed and altitude, the variance increased to 94%. For 304 users, the variance explained by heart-rate was 32.8% and increased to 65.9% with contextual features. This exploratory study highlights the potential of heart rate as a biometric identifier. It also underscores how contextual factors, such as speed and altitude change, can improve classification of timeseries data when coupled with smart data preprocessing techniques.

ABSTRACT. Predicting patients’ length of stay (LOS) is crucial for efficient scheduling of treatment and strategic future planning, in turn reduce hospitalisation costs. However, this is a complex problem requiring careful selection of optimal set of essential factors that significantly impact the accuracy and performance of LOS prediction. Using an inpatient dataset of 285k of records from 14 general care hospitals in Vermont, USA from 2013-2017, we presented our novel approach to incorporate novel features to improve the accuracy of LOS. Our empirical experiment and analysis showed considerable improvement in LOS prediction with an XGBoost model RMSE score of 6.98 and R2 score of 38.24%. Based on several experiments, we provided empirical analysis of the importance of different feature sets and its impact on predicting patients’ LOS.

ABSTRACT. Feature selection always constitutes a part of the knowledge discovery process. It can be based entirely on domain knowledge, dictated by algorithms that estimate attribute relevance, or combine the two approaches with the aim of finding the most important variables. The degrees of importance for a considered task can vary significantly, in particular when the input domain is continuous. To reflect the relevance of features, a ranking mechanism can be used. The paper demonstrates the research methodology focused on observations of relations between attribute relevance, displayed by rankings, and discretisation. Instead of transforming all continuous attributes before data exploration, the variables were gradually processed, and the impact of such a change on the performance of a classifier was studied. The considerable experiments performed on stylometric data illustrate that controlled and selective discretisation could be more advantageous to predictive accuracy than uniform transformation of all features.

ABSTRACT. Interactive visual exploration of large high-dimensional data (HDD) plays a very important role in various scientific fields that require aggregated information about the interrelationships between numerous objects. We argue that the visualization of very large HDD is well approximated by the two-dimensional problem of embedding undirected \textit{k}NN-graphs. In the advent of the big data era, the size of complex networks (datasets) $G(V,E)$ ($|V|$$=$$M$$\sim$$10^{6+}$) represents a great challenge for today's computer systems and still requires more efficient $ND$$ \rightarrow$$2D$ dimensionality reduction (DR) algorithms. Existing DR methods, which involve greater computational and memory complexity than $O(M)$, are too slow for interactive manipulation of large data that involve millions of samples. We investigate how the quality of HDD visualization and its computational complexity changes with successive simplifications of well known UMAP and t-SNE algorithms by reducing the number of nearest neighbors and the distance metric. We show that the radically simplified IVHD algorithm, although producing embeddings inferior to its competitors for the most of small and moderate data sizes ($M$$<$$10^5$ samples), is still amazingly effective. It reconstructs with sufficient precision both local and global data topology. However, it is much faster than state-of-the-art algorithms. IVHD shows its power for larger datasets ($M$$\sim$$10^6$) for which the baseline methods fail in a reasonable computational time.

ABSTRACT. Missing data can be described by the absence of values in a dataset, which can be a critical issue in domains such as healthcare. A common solution for this problem is imputation, where the missing values are replaced by estimations. Most imputation methods are suitable for the Missing Completely At Random (MCAR) and Missing At Random (MAR) mechanisms but produce biased results for Missing Not At Random (MNAR) values. An effective approach to mitigate this bias effect is to use the delta-adjustment method. This method assumes the imputation is performed for the MAR mechanism and adjusts the imputed values to become valid under MNAR assumptions by applying a correction factor. Such adjustment is usually defined manually by a domain expert, which often makes this method unfeasible. In this work, we propose an automatic procedure to find an approximate delta adjustment value for every feature of the dataset, which we call Automatic Delta-Adjustment Method. The proposed procedure is validated in an experimental setup comprising 10 datasets of the healthcare domain injected with MNAR values. The results from seven state-of-the-art imputation methods are compared with and without the adjustment, and applying the correction provides a significantly lower imputation error for all methods.

ABSTRACT. We consider a novel modeling and computational framework of the distribution-ally-robust optimization problem of jump-driven general affine process focusing on its application to inland fisheries as a case study. In particular, we consider an exploitation problem of migrating fish population as a major fishery issue. Our target process is the superposition of Ornstein–Uhlenbeck processes (supOU process) serving as a fundamental model of mean-reverting phenomena with (sub-)exponential memory. It is an affine process and admits a closed-form char-acteristic function being useful in applications. A theoretical novelty here is the assumption that the supOU process is allowed to have uncertain model parameter values as often encountered in engineering applications. Another novelty is the formulation of a long-term exploitation problem of the supOU process where the uncertainty is penalized through a generalized divergence between benchmark and distorted models. We present a strictly convex discretization of the optimization problem based on the model identified using the existing data of migrating fish population of a river in Japan. Further, the statistical analysis results in this paper are new by themselves. The computational results suggest the optimal harvesting policy of the fish population.

ABSTRACT. Estimation of approximation errors on an ensemble of numerical so-lutions obtained by independent algorithms is addressed in the linear and non-linear statements. In linear case the influence of the irremovable uncertainty on error estimates is considered. In nonlinear case, the nonuniform improve-ment of estimates’ accuracy is demonstrated that enables to overperform the quality of linear estimates. An ensemble of numerical results, obtained by four OpenFOAM solvers, is used as the input data. A comparison of approximation errors, obtained by these methods, and the exact error, computed as the differ-ence of numerical solutions and the analytical solution, is presented for the in-viscid compressible flow with an oblique shock wave. The numerical tests demonstrated feasibility to obtain the reliable error estimates (in linear case) and to improve the accuracy of certain approximation error in nonlinear case.

ABSTRACT. In this work, we introduce and study a set of tree-based algorithms for resources allocation considering group dependencies between their parameters. Real world distributed and high-performance computing systems often operate under conditions of the resources availability uncertainty caused by uncertainties of jobs execution, inaccuracies in runtime predictions and other global and local utilization events. In this way we can observe an availability over time function for each resource and use it as a scheduling parameter. As a single parallel job usually occupies a set of resources, they shape groups with common probabilities of usage and release events. The novelty of the proposed approach is an efficient algorithm considering groupings of resources by the common availability probability for the resources’ co-allocation. The proposed algorithm combines dynamic programming and greedy methods for the probability-based multiplicative knapsack problem with a tree-based branch and bounds approach. Simulation results and analysis are provided to compare different approaches, including greedy and brute force solution.

ABSTRACT. The two-dimensional scattered data interpolation problem is investigated. In contrast to the traditional Method of Radial Basis Functions, the interpolation problem is converted to a higher order (biharmonic or modified bi-Helmholtz) partial differential equation supplied with usual boundary conditions as well as pointwise interpolation conditions. To solve this fourth-order problem, the Method of Fundamental Solutions is used. The source points, which are needed in the method, are located partly in the exterior of the domain of the corresponding partial differential equation and partly in the interpolation points. This results in a linear system with possibly large and fully populated matrix. To make the computations more efficient, a localization technique is applied, which splits the original problem into a sequence of local problems. The system of local equations is solved in an iterative way, which mimics the classical overlapping Schwarz method. Thus, the problem of large and ill-conditioned matrices is completely avoided. The method is illustrated via a numerical example.

ABSTRACT. In this paper, we address a way to reduce the total computational cost of meshless approximation by reducing the required stencil size through spatial variation of computational node regularity. Rather than covering the entire domain with scattered nodes, only regions with geometric details are covered with scattered nodes, while the rest of the domain is discretized with regular nodes. Consequently, in regions covered with regular nodes the approximation using solely the monomial basis can be performed, effectively reducing the required stencil size compared to the approximation on scattered nodes where a set of polyharmonic splines is added to ensure convergent behaviour. The performance of the proposed hybrid scattered-regular approximation approach, in terms of computational efficiency and accuracy of the numerical solution, is studied on natural convection driven fluid flow problems. We start with the solution of the de Vahl Davis benchmark case, defined on square domain, and continue with two- and three-dimensional irregularly shaped domains. We show that the spatial variation of the two approximation methods can significantly reduce the computational complexity, with only a minor impact on the solution accuracy.

ABSTRACT. The paper discusses various formulations of the recently developed higher order multipoint meshless finite difference method. The novel multipoint approach is based on raising the order of approximation of the unknown function by introducing additional degrees of freedom in a stencil nodes, taking into account e.g. the right hand side of the considered differential equation. It improves the finite difference solution without increasing the number of nodes in an arbitrary irregular mesh. There two basic versions of the Multipoint MFDM – the general and specific one. Other extensions of the multipoint meshless FDM allow for analysis of boundary value problems posed in various weak formulations, including variational ones (Galerkin, Petrov-Galerkin), minimization of the energy functional, and MLPG. Several versions of the multipoint method are proposed and examined. The paper is illustrated with several examples of the multipoint numerical tests carried out for the weak formulations and their comparison with those obtained for the strong formulation.

ABSTRACT. The reconstruction of gap-free signals from observation data is a critical challenge for numerous application domains, such as geoscience and space-based earth observation, when the available sensors or the data collection processes lead to irregularly-sampled and noisy observations. Optimal interpolation (OI), also referred to as kriging, provides a theoretical framework to solve interpolation problems for Gaussian processes (GP). The associated computational complexity being rapidly intractable for n-dimensional tensors and increasing numbers of observations, a rich literature has emerged to address this issue using ensemble methods, sparse schemes or iterative approaches. Here, we introduce a neural OI scheme. It exploits a variational formulation with convolutional auto-encoders and a trainable iterative gradient-based solver. Theoretically equivalent to the OI formulation, the trainable solver asymptotically converges to the OI solution when dealing with both stationary and non-stationary linear spatio-temporal GPs. Through a bi-level optimization formulation, we relate the learning step and the selection of the training loss to the theoretical properties of the OI, which is an unbiased estimator with minimal error variance. Numerical experiments for 2D+t synthetic GP datasets demonstrate the relevance of the proposed scheme to learn computationally-efficient and scalable OI models and solvers from data. As illustrated for a real-world interpolation problems for satellite-derived geophysical dynamics, the proposed framework also extends to non-linear and multimodal interpolation problems and significantly outperforms state-of-the-art interpolation methods, when dealing with very high missing data rates.

ABSTRACT. Data assimilation methods are mainly based on the Bayesian formulation of the estimation problem. For cost and feasibility reasons, this formulation is usually approximated by Gaussian assumptions on the distribution of model variables, observations and errors. However, when these assumptions are not valid, this can lead to non-convergence or instability of data assimilation methods. The work presented here introduces the use of kernel methods in data assimilation to model uncertainties in the data in a more flexible way than with Gaussian assumptions. The use of kernel functions allows to describe non-linear relationships between variables. The aim is to extend the assimilation methods to problems where they are currently unefficient. The Ensemble Transform Kalman Filter (ETKF) formulation of the assimilation problem is reformulated using kernels and show the equivalence of the two formulations for the linear kernel. Numerical results on the toy model Lorenz 63 are provided for the linear and hyperbolic tangent kernels and compared to the results obtained by the ETKF.

ABSTRACT. We introduce a gradient-free data-driven framework for optimizing the power output of a wind farm based on a Bayesian approach and large-eddy simulations. In contrast with conventional wind farm layout optimization strategies, which make use of simple wake models, the proposed framework accounts for complex flow phenomena such as wake meandering, local speed-ups and the interaction of the wind turbines with the atmospheric flow. The capabilities of the framework are demonstrated for the case of a small wind farm consisting of five wind turbines. It is shown that it can find optimal designs within a few iterations, while leveraging the above phenomena to deliver increased wind farm performance.

ABSTRACT. With: L. Bertino, M. Bocquet, J. Brajard, Y. Chen, S. Driscoll, C. Durand, A. Farchi, T. Finn, C. Jones and I. Pasmans

We present an overview of the research efforts and results obtained in the context of the international project SASIP aimed at understanding and prediction the Arctic changes. We have been working on developing novel data assimilation, machine learning and their combination adapted to a new generation of sea-ice models that treats the ice as a brittle solid instead of as a fluid. These models present unique physical challenges such as sharp gradients, anisotropy and multifractality. In this talk we will first present the application of an ensemble variational method to estimate the state and parameters of the sea-ice model based on synthetic, satellite-like, data, illustrating the power and limitation of the available measurements. Second, we will show how to adapt the data assimilation procedure to the use of discontinuous Galerkin model, a modification that makes possible to assimilate very dense data (such as satellite) as well as to develop a scale-aware localisation procedure. To incorporate multifractal, anisotropic, and stochastic-like processes in sea ice, we envision the combination of geophysical sea-ice models together with neural networks in a hybrid modelling setup. On the one hand, deep learning can surrogate computationally expensive sea-ice models, on the other hand, deep learning can parametrize subgrid-scale processes in sea-ice models and correct persisting model errors, improving the forecasts by up to 70 % across all model variables on an hourly timescale. Finally, we will show how to use neural networks to emulate and replace a physical parametrization of the sea-ice melt ponds that create on the ice surface, and that have a major role on the albedo and thus on the general energy balance. Overall, our results show the potential of data assimilation and machine learning to extract much information from available data to correct model prediction and the models themselves.

ABSTRACT. This paper presents a method for the tracking of multiple drones in three-dimensional space based on data from a multi-camera system. It uses the Particle Swarm Optimization (PSO) algorithm and methods for background/foreground detection. In order to evaluate the developed tracking algorithm, the dataset consisting of three simulation sequences and two real ones was prepared. The sequences contain from one to ten drones moving with different flight patterns. The simulation sequences were created using the Unreal Engine and the AirSim plugin, whereas the real sequences were registered in the Human Motion Lab at the Polish-Japanese Academy of Information Technology. The lab is equipped with the Vicon motion capture system, which was used to acquire ground truth data. The conducted experiments show the high efficiency and accuracy of the proposed method. For the simulation data, tracking errors from 0.08m to 0.20m were obtained, while for real data, the error was 0.11-0.13m. The system was developed for augmented reality applications, especially games. The dataset is available at http://bytom.pja.edu.pl/drones/.

ABSTRACT. We propose a method for retrieving solar magnetograms based on their content. We leverage data collected by the SDO Helioseismic and Magnetic Imager using SunPy and PyTorch libraries to create a vector-based mathematical representation of the Sun's magnetic field regions. This approach enables us to compare short vectors instead of comparing full-disk images. To reduce retrieval time, we use a fully-connected autoencoder to compress the 144-element descriptor to a 36-element semantic hash. Our experimental results demonstrate the efficiency of our approach, which achieved the highest precision value compared to other state-of-the-art methods. Our proposed method is not only applicable for solar image retrieval but also for classification tasks.

ABSTRACT. Reverse engineering is one of the classical approaches for quailty assessment in industrial manufacturing. A key technology in reverse engineering is surface reconstruction, which aims at obtaining a digital model of a physical object from a cloud of 3D data points obtained by scanning the object. In this paper we address the surface reconstruction problem for surfaces that can exhibit large smooth bumps. To account for this type of features, our approach is based on using exponentials of polynomial functions in two variables as the approximating functions. In particular, we consider three different models, given by bivariate distributions obtained by combining a normal univariate distribution with a normal, Gamma, and Weibull distribution, respectively. The resulting surfaces depend on some parameters whose values have to be optimized. This yields a difficult nonlinear continuous optimization problem solved through an artificial immune systems approach based on the clonal selection theory. The performance of the method is discussed through its application to a benchmark comprised of three examples of point clouds.

ABSTRACT. This paper presents methodology of prediction of casting mechanical parameters based on direct microstructure image analysis using deep neural networks and graphite forms recognition and classification. These methods are applied to predict tensile strength of iron-carbon alloys based on microstructure photos taken with the light-optical microscopy technique but are general and can be adapted to other applications. In the first approach EfficientNet architecture is used. In the second approach graphite structures are separated, recognized using VGG19 network, counted and classified using support vector machines, decision trees, random forest, logistic regression, multi-layer perceptron or AdaBoost. Performance of the first approach was better. However, the second allowed to create classifier which can be easily analyzed by human expert.

ABSTRACT. Today, android devices like smartphones, tablets, etc., have penetrated very deep into our modern society and have become an integral part of our daily lives. The widespread adoption of these devices has also garnered the immense attention of malware designers. Many recent reports suggest that existing malware detection systems cannot cope with current malware challenges and thus threaten the android ecosystem's stability and security. Therefore, researchers are now turning towards android malware detection systems based on machine and deep learning algorithms, which have shown promising results. Despite their superior performance, these systems are not immune to adversarial attacks, highlighting a research gap in this field. Therefore, we design and develop EvoAAttack based on a genetic algorithm to expose vulnerabilities in state-of-the-art malware detection systems. The EvoAAttack is a targeted false-negative evasion attack strategy for the grey-box scenario. The EvoAAttack aims to convert malicious android applications (by adding perturbations) into adversarial applications that can deceive/evade detection systems. The EvoAAttack agent is designed to convert maximum malware into adversarial applications with minimum perturbations while maintaining syntactic, semantic, and behavioral integrity. We tested EvoAAttack against thirteen distinct malware detection systems based on machine and deep learning algorithms from four different categories (machine learning, bagging, boosting, and neural networks). The EvoAAttack was able to convert an average of 97.48% of malware applications (with a maximum of five perturbations) into adversarial applications (malware variants). These adversarial applications force misclassifications and reduce the average accuracy of thirteen malware detection systems from 94.87% to 50.31%. Later we also designed a defense strategy (defPCA) to counter the adversarial attacks. The defPCA defense reduces the average forced misclassification rate from 97.48% to 59.98% against the same thirteen malware detection systems. Finally, we conclude that threat modeling improves both detection performance and adversarial robustness of malware detection systems.

ABSTRACT. Specific research programs, legislation and funding intend to protect, con-serve and enhance the EU's natural capital, transforming the EU into a green, competitive, low-carbon and resource-efficient economy. These guidelines aim at protecting European citizens from health and environmental risks. In-deed, there is an increasing interest on decarbonization of the electricity gen-eration, with a special focus on the introduction of Renewable Energy Re-sources (RES). Solar energy, a form of renewable energy, is not only is abundant in mediterranean countries but can reduce the harmful CO2 emis-sions that result from the burning of fossil fuels. The most common way of using solar power is to convert sunlight into heat energy to produce hot wa-ter, through the usage of solar thermal collectors. This paper is a preliminary insight into a new control approach from where smart decision is made based on predictions returned by models of sustainable thermal systems (lo-cal renewable sources generation devices) and on information gathered from an array of sensors in order to regulate swimming pool’s water temperature. The information (ambient variables and sub-systems internal transfer func-tion modelling) is then combined with an optimization framework which goal is to ultimately, reduce the requirements for human intervention in the swimming pool maintenance and provide resource savings for the final user, in terms of financial and natural resources, contributing to a sustainable en-vironment. The research work is developed within the scope of the Eco-pool+++ project: Innovative heated pools with reduced thermal losses.

ABSTRACT. In an industrial setting, predicting the remaining useful life- time of equipment and systems is crucial for ensuring efficient operation, reducing downtime, and prolonging the life of costly assets. There are state-of-the-art machine learning methods supporting this task. How- ever, in this paper, we argue, that both efficiency and understandability can be improved by the use of explainable AI methods that analyze the importance of features used by the machine learning model. In the paper, we analyze the feature importance before a failure occurs to identify events in which an increase in importance can be observed and based on that indicate attributes with the most influence on the failure. We demonstrate how the analyses of Shap values near the occurrence of failures can help identify the specific features that led to the failure. This in turn can help in identifying the root cause of the problem and developing strategies to prevent future failures. Additionally, it can be used to identify ar- eas where maintenance or replacement is needed to prevent failure and prolong the useful life of a system.

ABSTRACT. The nonmetallic pipes can effectively avoid the hydrogen embrittlement of metal pipes when transporting hydrogen. However, due to the characters of the nonmetal materials, there will be a large degree of gas permeation and leakage when conveying hydrogen by nonmetallic pipes. To select suitable nonmetal pipe materials, the solubility, diffusion and permeability of hydro-gen in PE, PVC and PVDF amorphous polymers are investigated and com-pared by molecular dynamics simulation at 270~310 K and 0.1~0.7 MPa, providing guidance for the construction of nonmetallic hydrogen transportation pipes. Simulation results indicate that the solubility coefficients of hydrogen in PE and PVDF rise with the increasing temperature, but show an opposite trend in PVC. Both the diffusion and permeability coefficients in-crease with the rise of temperature. In a small range of pressure variation, the influence of pressure on diffusion and permeation characteristics is ignorable. Among the three studied amorphous polymers, the permeability coefficient of H2 in PE is the largest and that in PVDF is the smallest. The diffusion of H2 molecules in the polymer conforms to the hopping mechanism that H2 molecules vibrate in pores with long time and hop to other pores with short time.

ABSTRACT. A novel unstructured conservative level-set method for film boiling is introduced. Transport equations are discretized by the finite-volume method on collocated unstructured grids. Mass transfer by thermal phase change is calculated using the temperature gradient in the liquid and vapour phases at the interface. The fractional-step projection method is used for solving the pressure-velocity coupling. The convective term of transport equations is discretized by unstructured flux-limiter schemes to avoid numerical oscillations around the interface and minimize numerical diffusion. The central difference scheme discretizes diffusive terms. Verification and validation for film boiling on a flat surface are performed.

ABSTRACT. Finite element (FE) simulation is an established approach to mechanical simulation of angioplasty and stent deployment. Agent-based (AB) models are an alternative approach to biological tissue modeling in which individual cells can be represented as agents and are suitable for studying biological responses. By combining these two approaches, it is possible to leverage the strengths of each to improve the in silico description of angioplasty or stenting and the following healing response. Here we propose a couping between FE and AB vascular tissue models using trilinear interpolation, where the stresses (and strains) in the AB model arise directly from the forces of interaction between individual agents. The stress values for FE and AB models are calculated and compared.

ABSTRACT. During the last years, a wide variety of epidemic models was employed to analyze the spread of COVID-19. As a rule, the modeling aim is to assess the impact of the epidemics and the efficiency of control measures to reduce the social and economic toll. In this regard, finding the most suitable model according to the available epidemic data is an important task to consider. In this project, we aim to perform a comparison of several approaches related to COVID-19 propagation modeling and analyze the dependence of their accuracy on their structural complexity, using COVID-19 dynamics data in St. Petersburg in 2020-2022. Our assessment is based on Akaike information criterion (AIC), which is widely used for the same purpose in statistics and machine learning, although its usage in explanatory modeling is somewhat limited. The results of the study help understand how to assess the complexity of explanatory models required to achieve the most accurate replication of disease outbreaks for epidemic data with a certain level of detail.

ABSTRACT. Epidemics of acute respiratory infections such as influenza and COVID-19 pose a serious threat to public health. To control the spread of infections, it is necessary to estimate epidemic indicators via statistical methods and mathematical models. When calculating these indicators, uncertainty arises, associated with bias in the initial data for model calibration. The dependence of modeling on the accuracy of the data can be huge, and the lack of its consideration, which is typical for most of the works on modeling the spread of epidemic ARIs, makes it difficult to correctly predict the effectiveness of anti-epidemic measures. In this research, we present methods and algorithms for calculating the uncertainty in estimating epidemic indicators when modeling the dynamics of epidemic ARI. We show the application of the methods for the case of an influenza outbreak in St. Petersburg.

ABSTRACT. Research projects working with Tuberculosis clinical data generate large volumes of complex data, requiring sophisticated tools to create processing pipelines to extract meaningful insights. However, creating this type of tool is a complex and costly task, especially for researchers who need to gain experience with technology or statistical analysis. In this work, we present a web portal that can connect to any database, providing easy access to statistical analyses of the clinical data in real time using charts, tables, or any other data visualization technique. The tool is user-friendly and customizable, reaching the project's needs according to its particularities. The developed portal in this work was used as a use case for the research project developed by the Federal University of Rio de Janeiro (UFRJ) for the validation and cost of performance of the Line Probe Assay 1 and 2 (LPA) as a method of diagnosing resistant Tuberculosis in Brazilian's reference centers. In the use case, the tool proved to be a valuable resource for researchers, bringing efficiency and effectiveness in analyzing results for quick and correct clinical data interpretation.

ABSTRACT. The use of health data in research is fundamental to improve health care, health systems and public health policies. However, due to the intrinsic sensitivity of this type of data, there are several privacy and security concerns that must be addressed to comply with best practices recommendations and the legal framework. Decentralized-learning methods, such as split- and federated-learning, allow the training of algorithms across multiple locations without data sharing. The application of those methods to the medical field holds great potential due to the guarantee of data privacy compliance. In this study, we performed a bibliometric analysis to explore the intellectual structure of this new research field and to assess its publication to collaboration patterns. A total of 3023 unique documents published between 2013 and 2023 were retrieved from Scopus, from which 488 were included in this review. The most frequent publication source was the IEEE Journal of Biomedical and Health Informatics (n=27). China was the country with the highest number of publications in this field, followed by USA. The top three authors were Dekker A (n=14), Wang X (n=13), Li X (n=12). The most frequent keywords were “Federated learning” (n=218), followed by “Deep learning” (n=62) and “Machine learning” (n=52). This study provides an overall picture of the research literature regarding the application of decentralized-learning in healthcare, possibly setting ground for future collaborations.