Precise Language Deception: XAI Driven Targeted Adversarial Examples with Restricted Knowledge
ABSTRACT. In this paper, we propose a novel approach for crafting targeted adversarial examples (attacks) using explainable artificial intelligence (XAI) techniques. Our method leverages XAI to identify key input elements that, when altered, can mislead NLP models, such as BERT and large language models (LLMs), into producing specific incorrect outputs. We demonstrate the effectiveness of our targeted attacks across a range of NLP tasks and models, even in scenarios where internal model access is restricted. Our approach is particularly effective in zero-shot learning settings, underscoring its adaptability and transferability to both traditional and conversational AI systems. In addition, we outline mitigation strategies, demonstrating that adversarial training and fine-tuning can enhance model defenses against such attacks. Although our work highlights the vulnerabilities of LLMs and BERT models to adversarial manipulation, it also lays the groundwork for developing more robust models, advancing the dual goal of understanding and securing black-box NLP systems. Through targeted adversarial examples and SHAP-based techniques, we not only expose the weaknesses of existing models but also propose strategies to enhance AI's resilience to deceptive linguistic input.
Towards weight-space interpretation of Low-Rank Adapters for Diffusion Models
ABSTRACT. Low-rank adapters (LoRAs) have emerged as an efficient method for customizing large-scale diffusion models, but their internal representations remain poorly understood. We present a comprehensive investigation of the interpretability of adapter weight-spaces for image diffusion models. To that end, we open-source a dataset of 100,000 Stable Diffusion adapters fine-tuned across a hierarchy of image concepts amounting to 264 leaf classes, complete with training metadata. Through systematic analysis, we demonstrate that adapter weights encode meaningful semantic information about their training data, enabling direct interpretation without image generation. We evaluate multiple weight-space representations, including raw parameters, statistical summaries, and learned embeddings, to determine their effectiveness in predicting training data characteristics. To demonstrate real-world impact, we apply our findings to the critical task of detecting potentially harmful content on newly introduced NSFW (Not Safe For Work) toy dataset of Stable Diffusion LoRAs fine-tuned on harmful content. This work advances the interpretability of adapter-based fine-tuning and provides practical tools for understanding and auditing adapted diffusion models.
Dynamic neural network with matrix-extended residual connections
ABSTRACT. The issue of adjusting neural network structure is one of the core problems in artificial intelligence. A highly desirable scenario is that of a dynamic model when structural changes occur while the model is training. In this paper, we propose a new and powerful tool that facilitates dynamic changes in network structure. We introduce residual connections based on matrix extensions. This novel mechanism allows for the adjustment of weight matrices to enhance the potential for structural modifications. We conducted a series of comprehensive experiments confirming that the new residual connections scheme behaves very well. The new type of connection improves performance by enabling better error flow during the error backpropagation phase, resulting in more efficient training. Our method demonstrates superior performance and enhanced trackability during the training process. The paper is supplemented by Python source code to ensure reproducibility. This method marks a significant starting point, showing immense potential for more advanced dynamic neural network models and transfer learning with dynamic models.
An Empirical Assessment of LLM-Based Approaches to Malicious Webpage Detection
ABSTRACT. Large language models (LLMs) are increasingly influential in advancing NLP technology and solving complex tasks, yet their potential misuse in cybersecurity poses significant risks. This paper addresses the challenge of detecting malicious webpages using LLMs, an area with limited research. We evaluate LLMs by expanding zero-shot and few-shot query formulations, testing previously unassessed open-source and proprietary models, and assessing robustness under adversarial conditions. Additionally, we verify model performance using Chain of Thought reasoning and compare these explanations with traditional methods. Our work aims to enhance the application of LLMs in cybersecurity, guiding the development of more effective detection systems.
Uncovering and Verifying Optimal Community Structure in Complex Networks: A MaxSAT Approach
ABSTRACT. Network modularity is central to understanding phenomena in diverse domains, from biology and social science to engineering and computational physics. However, computing the optimal modularity—an NP-hard measure quantifying community strength—has remained computationally intractable at large scales. Most approaches resort to heuristics without formal optimality guarantees.
This paper contributes to the computational science of complex systems by introducing a novel MaxSAT-based framework that can compute optimal network modularity values for larger networks than previously possible. Leveraging this new capability, we extensively evaluate heuristic solutions and, for the first time, include the state-of-the-art memetic graph clustering heuristic VieClus. Remarkably, VieClus identifies optimal modularity values for all tested networks, ranging from 103 previously studied instances to 52 new, larger ones, and does so in seconds. This result contrasts with earlier conclusions that heuristics frequently fail to find the optimal modularity.
By combining a powerful MaxSAT encoding, which supports proof logging for verification, with a fast and effective heuristic, we demonstrate that even intricate network structures can be tackled efficiently. This synergy brings us closer to making complex network analysis and community detection tractable, robust, and verifiable—a goal firmly aligned with the core mission of computational science.
cuTeBool: Fast and Scalable Boolean matrix factorization on GPUs using Tensor Cores
ABSTRACT. Boolean matrix factorization aims to represent binary data as a product of two factor matrices, in order to uncover the underlying structure of the data and find a compressed representation. However, finding the factors of a given ground truth is computationally hard and calls for fast implementations that accomplish a good approximation in reasonable time.
We present cuTeBool, a novel parallel algorithm that exploits Tensor Cores on CUDA-enabled GPUs for fast matrix operations based on a randomized approach. Our comprehensive performance evaluation shows that it produces approximate factorization competitive to other state-of-the-art tools within vastly reduced runtime for a variety of input matrices. Moreover, our algorithm is the only available method that scales well with the size of the ground truth and is able to factorize matrices that are at least one order-of-magnitude larger than all competitors. We further analyze algorithmic parameters allowing us to find a trade-off between performance and reconstruction quality.
Dimensionality reduction in product of metric spaces
ABSTRACT. The purpose of the article is to develop a new dimensionality reduction algorithm for data that are describes by many features of different nature. A method of feature selection is based on a new concept of metrical importance of the features. Our concept of the feature importance is based on metrical properties of data and is inspired by the principle component analysis. Numerical experiments confirm the effectiveness of the method and certain accordance of it with other concepts of feature importance.
Hierarchical Structural Information -- Theory and Applications
ABSTRACT. This paper introduces a novel measure to quantify structural information in hierarchical graphs. It addresses the limitation of current methods that do not adequately account for hierarchical structures. By considering inner structural information and distinguishability of higher-level vertices, the proposed measure captures the additional information generated by the hierarchy. The hypothesis that hierarchical graphs contain more structural information is validated using the "Countries" dataset. Results demonstrate a measurable increase in information content when the hierarchical structure is considered, compared to a simple graph representation. This highlights the importance of recognizing and utilizing hierarchy to enhance the informational richness of graphs, potentially improving the performance of graph-based machine learning models.
Proof of Training: Obtaining Verifiable ML Models by Delegating Training to a Blockchain Network
ABSTRACT. The recent rise of Bitcoin has sparked an unprecedented trend of enthusiasts acquiring expensive hardware for mining. This self-perpetuating race, driven by Bitcoin’s PoW consensus mechanism, has led to the creation of massive computational centers dedicated solely to solving impractical hash inversions. However, these computational resources could be redirected toward more meaningful tasks if nodes were properly incentivized.
In this paper, we introduce Proof of Training (PoT), a novel consensus mechanism that offers two key advantages over previous approaches. First, it replaces wasteful computations with ML training. Second, by aligning the inherent distrust between nodes in blockchain networks with distributed model training, PoT not only achieves consensus but also produces a trained model along with proof that it was trained on the client-provided data.
PoT enables clients to hire the blockchain network to provably train arbitrary models using their provided datasets and architectures. Meanwhile, nodes that participate in training are rewarded with PoT cryptocurrency based on their computational contributions.
MinRNNs for Lagrangian-Based Simulations of Transient Flow Problems
ABSTRACT. Motivated by the need for faster yet accurate surrogate modeling of continuum simulations, we investigate whether the recently proposed minimal recurrent networks (minLSTM and minGRU) can benefit particle-based fluid and soft-solid simulations. To our knowledge, this is the first work applying these minimal RNNs to Lagrangian data from 2D continuum simulation, including single-phase fluids and multimaterial interactions. We embed minLSTM and minGRU in an MLP based
encoder–decoder and compare them against (i) a classical LSTM, and (ii) an MLP baseline with no recurrent core. Where prior studies of minRNNs focused on simpler time-series tasks, our results show that minLSTM and minGRU remain highly effective in these physics-driven settings: they train approximately 350–400% faster than the standard
LSTM or GRU, while matching—and often surpassing—their accuracy. Thus, for particle-based continuum simulations, minimal recurrent architectures offer a superior trade-off between computational overhead and predictive performance, thereby advancing real-time or high-fidelity simulation workflows in engineering and visual effects. We conclude that minimal RNNs are well-suited for surrogate modeling of fluid and softsolid dynamics.
ABSTRACT. Geosteering, the art of navigating wells to maximize the reservoir resources, is fraught with challenges of geological uncertainty and the relentless pace of real-time operations. In this paper, we present a novel framework that integrates Particle Filters (PF) for probabilistic subsurface interpretation with a Dual-Network Deep Reinforcement Learning (DRL) model for adaptive decision-making in geosteering operations.
The PF component quantifies subsurface uncertainties, providing a probabilistic interpretation of geological boundaries, while the DRL model leverages this information to generate optimal steering decisions. This synergy ensures robust trajectory planning that dynamically adapts to real-time geological changes. The framework incorporates key features, such as target-line alignment to maintain wellbore proximity to reservoir zones and dog-leg severity constraints to ensure operational feasibility.
Extensive verification in an industry-standard environment accessed via an API demonstrates the model’s ability to accurately track reservoir boundaries, predict gamma-ray values, and optimize well trajectories. The results highlight significant improvements over traditional geosteering approaches and standard DRL-based methods in terms of reservoir contact, decision-making efficiency, and trajectory accuracy, even in low-data scenarios. The proposed framework provides a scalable and robust solution for quantifying uncertainties in real-time geosteering, paving the way for informed operational decisions improving value-creation and drilling effciency.
Performance-energy investigation of selected applications using a parallel multi-GPU genetic algorithm under power capping
ABSTRACT. In this article we demonstrate performance-energy optimization of multi-GPU genetic algorithm execution using power capping.
Firstly, we outline elements concerning design and implementation of a multi-GPU framework for execution of a genetic algorithm, allowing to apply a solution to a variety of problems. Secondly, implementation of three algorithms is proposed and discussed: Traveling Salesman, Knapsack and Partition. Then, we present a testbed environment with a high performance computing node with 2 multi-core Intel Xeon CPUs and 8 NVIDIA Quadro RTX 6000 GPUs as well as a Yokogawa WT-310E power meter. Finally, we describe and discuss results of optimization of the implementations using 1, 2, 4 and 8 GPUs under different power caps imposed by NVIDIA NVML. We show scalability of the solution in terms of fitness versus the number of GPUs used and analyze execution times and energy consumption of various cases under various power caps. We demonstrate that for 8 GPUs, using the power cap of 140W per GPU, we can obtain considerable energy savings of
over 17.93% for Traveling Salesman,
15.88% for Knapsack,
21.97% for Partition,
with small
increases of execution time: 0.89% for Traveling Salesman, 1.41% for Knapsack and 14.64% for Partition, versus the results for the default power cap of 260W per GPU.
Discrete Residual Loss Functions for Training Physics-Informed Neural Networks
ABSTRACT. The use of neural networks and operators to solve partial differential equations that govern fluid flow is carried out in a simulation-free, physics-informed approach, where the residual of the governing equation calculated via automatic differentiation across the neural network is the loss function used for training. One issue with this approach is that simulating highly non-linear flows such as high Reynolds number flows is challenging, even in laminar settings. We propose a new simulation-free approach for training PINNs using a residual loss based on a discrete numerical scheme instead of automatic differentiation. This loss function also requires a new grid-based PINNs training strategy. Using the loss landscape, we demonstrate why our new loss function works better than the automatic differentiation-based loss function. We also demonstrate how to implement our grid-based training for complex geometry. Simulations using the new neural model for high Reynolds number fluid flow and complex geometry test cases are showcased and compared with automatic differentiation approaches. The results show that our new discrete loss function and training strategy take less computational time, converge faster than automatic differentiation, and can be used to simulate non-linear flow efficiently.
Graph-augmented Large Language Models Can Reach Human Intelligence in Clinical Applications
ABSTRACT. While recent advances in large language models (LLMs) offer great promise in healthcare, the safety-critical nature of the domain requires a thoughtful strategy to mitigate risks of hallucinations and potential harms. We propose and empirically validated that a graph of domain knowledge, injected as a source of truth, presents a promising approach to safe, factual, and in turn more precise LLM applications. We first introduce the world's first clinical terminology for the Chinese healthcare community, namely MedCT, accompanied by a clinical foundation model MedBERT and an entity linking model MedLink. The MedCT system enables standardized representation of clinical data, successively stimulating the development of new medicines, treatment pathways, and better patient outcomes. Moreover, the MedCT knowledge graph provides a principled mechanism to minimize the hallucination problem of LLMs, therefore achieving significant levels of accuracy and safety in LLM-based clinical applications. Our experiments show that the MedCT system achieves state-of-the-art (SOTA) performance in semantic matching and entity linking tasks, not only for Chinese but also for English. We also conducted a longitudinal field experiment by applying MedCT and LLMs in a representative spectrum of clinical tasks, including electronic health record (EHR) auto-generation and medical document search. Our study shows a multitude of values of MedCT for clinical workflows and patient outcomes, especially in the new genre of clinical LLM applications. We present our approach in sufficient engineering detail, such that implementing a clinical terminology for other non-English societies should be readily reproducible. With our hope to motivate further research on LLM-based healthcare digitalization, and at large the wellbeing of humankind, we openly release our terminology, models and algorithms, along with real-world clinical datasets for the development.
Cross-Scale Modeling of Healthcare Norms and Patient Features Dynamics with Interpretable Machine Learning
ABSTRACT. This study proposes an interpretable machine learning framework to model bidirectional dynamic interactions between macroscopic norms and microscopic features in clinical data. Leveraging real-world medical records from a specialized chest hospital (containing unstructured text, complex categorical variables, temporal indicators, and nonrandom missing patterns), we perform numerical processing through Latent Semantic Analysis and dimensionality reduction via Non-negative Matrix Factorization. Macroscopic therapeutic norms are identified using HDBSCAN clustering, while SHAP-XGBoost integration selects critical microscopic features, including multidrug-resistant tuberculosis diagnosis and liver function biomarkers. We integrate symbolic regression with the Peter-Clark Momentary Conditional Independencecausal discovery method based on partial correlation, constructing cross-scale functional relationships with temporally rigorous constraints. Specifically, PySR derives nonlinear mapping equations, while partial correlationbased conditional independence tests establish time-lagged dynamic dependency networks. Guided by the Dynamic Maximum Entropy across Scales (DyMES) principle, multi-scale perturbation experiments reveal bidirectional mechanisms. Within our dataset and framework, DyMES reveals dynamic constraints’ interplay driving statistical equilibrium between macroscopic clinical norms and microscopic patient characteristics through nonlinear coordination and threshold-triggered time-encoded mechanisms. Persistent constraint interactions induce novel steady states formation with dynamically preserved system memory.
A Fractional Computation Based Deep Learning Framework for Silicosis Detection
ABSTRACT. This study presents a new fractional computational approach applied to a new dataset, silicosis. This is a scalable and flexible approach for training neural networks using fractional computation, which conveniently use the conformable fractional derivative. During the training process, the method includes an independent variable, α, which provides additional degree to the framework. Fractional variants of the sigmoid and relu activation functions are explored and compared to conventional activation functions. This method builds on earlier approaches by employing the conformable fractional derivative. The fractional activation functions notably converge to the actions of their standard version when α =1, guaranteeing a smooth integration with conventional neural network models. The study also tackles the problem of managing both positive and negative inputs, which is a crucial prerequisite for the derivative but has been mainly disregarded in earlier studies, underscoring the originality of the current work. The experimental framework incorporates both feedforward neural network and convolutional neural network using fractional activation functions. The findings indicate that the suggested framework performs better and is more accurate for particular values of α. The efficiency of the suggested computational approach is demonstrated by showing that fractional activation on Convolutional Neural Network when paired with transfer learning, performs better for silicosis chest X-ray classification than conventional transfer learning models.
Combining XAI and graph cuts for skin-lesion segmentation
ABSTRACT. Deep neural networks and supervised machine learning for segmenting medical images, including dermatology, require large pixel-wise annotated data sets for training, which can be difficult to obtain. Image classification however only requires a label for each image, which is often automatically provided with a medical diagnosis. Explainable-AI (XAI) algorithms provide a means to identify pixels in the original image that are part of the object or relevant structure. Our method exploits this information by constructing a network graph from XAI explanations and segmenting the image using the graph-cut algorithm. We evaluate our approach using the HAM10k dataset and show that it allows segmenting skin lesions in dermatoscopic images without using pixel-wise annotated data for training. This makes our approach a cost-efficient alternative in scenarios where no annotated images are available.
Automatic Detection and Segmentation of Coronary Artery Stenosis in Coronary Angiography Images
ABSTRACT. In this paper, we present an approach for the detection, segmentation, and quantification of stenoses in coronary arteries using modern computer vision and deep learning techniques. Our system incorporates a detection model based on YOLOv8 and a segmentation model (DeepLabV3+) for precise localization and delineation of stenosis regions. In addition, a novel method is introduced to measure arterial thickness to support clinical decision-making. The experimental evaluation shows that the approach demonstrates high quality and performance in comparison to existing solutions. This work aims to improve diagnostic efficiency and reduce the reliance on expensive foreign-made equipment by providing an integrated solution that can operate on standard hardware.
Numerical Analysis of Dolphin Kick in Competitive Swimming with Free Surface Effects
ABSTRACT. This study develops a high-precision numerical simulation method that considers free surface deformation to analyze fluid dynamics in dolphin kick swimming. We integrate an interface tracking method with the Moving Computational Domain (MCD) method and the unstructured moving-grid finite-volume method. The simulation validation demonstrates agreement with theoretical solutions, confirming the method's accuracy in analyzing flow fields with free surfaces. Using this approach, we analyze flow fields around dolphin kick swimmers and evaluate how gravity and free surface deformation affect propulsion and drag forces. The results indicate that the free surface experiences an upward displacement in front of the swimmer, a downward displacement above the back, and wave formation in the wake. Comparative analyses indicate that pressure distribution variations caused by gravity and vortex structures significantly influence propulsion and drag. Through simulations with varied joint angles and kick frequencies, we establish that swimming speed increases linearly with both joint angle amplification and stroke period reduction. Our analysis shows a proportional relationship between swimming velocity and the combined effect of increased joint angles and decreased kick cycle times. Our findings validate the effectiveness of the proposed simulation technique for determining optimal joint angles and stroke periods to achieve efficient swimming speeds in competitive dolphin kick swimming.
Robust, Efficient, and Long-Time Accurate Schemes to Simulate Gas Storage in Geological Formation
ABSTRACT. In this paper, we consider the numerical simulation of gas storage in geological formation in the context of hydrogen underground storage and carbon dioxide geological sequestration. We constructs two energy-stable numerical schemes: one based on the energy factorization approach, which rigorously preserves the energy dissipation principle and combines discontinuous Galerkin approximations with mixed finite elements for spatial discretization; the other based on a stabilization approach, which conserves the original energy functional, has an adaptive stabilization parameter and time-stepping strategy, and ensures the boundedness of molar density. Through numerical experiments with methane gas, our schemes are validated in terms of capturing coupled hydro-mechanical processes, handling strong nonlinearities, and maintaining conservation properties.
Massively Parallel Computational Modeling of Porous Electrode Formation and Performance via Non-Solvent Induced Phase Separation for Redox Flow Batteries
ABSTRACT. Renewable energy sources such as solar and wind are essential for mitigating climate change and reducing reliance on fossil fuels. However, their intermittent and unpredictable nature can lead to instability in the power grid. To address this issue, electrochemical energy storage technologies, such as Redox Flow Batteries (RFBs), have gained significant attention due to their scalability, high energy efficiency, safety, and environmental benefits, making them a promising solution for large-scale energy storage. At the core of these systems, the porous electrode is the critical component responsible for multiple functions in the cell, such as providing surface area for electrochemical reactions, an open structure for fluid flow and mass transport, and the conductive scaffold for electronic and thermal conduction.
Current electrode materials are primarily limited to carbon-fiber-based structures, with fabrication methods that restrict performance due to limited control over microstructure and surface properties. To overcome this, new techniques like non-solvent induced phase separation (NIPS) have been recently introduced, offering precise control over electrode morphology, including porosity gradients and multimodal porosity, by adjusting factors like polymer concentration, solvent type, and temperature. Tuning these parameters experimentally is mostly done by trial and error by levering intuition on polymer science and engineering, but this a challenging and resource-intensive task. On the contrary, computational modeling can help guide the design process by correlating the influence of polymer solution parameters on the resulting electrode morphology and ultimately on the electrochemical performance.
In this research, we have developed a high-performance framework for simulating the formation of NIPS electrodes and the pore-scale transport inside them. The formation model is based on phase-field modeling of the phase separation process developed using the finite difference method implemented using NVIDIA CUDA to run on GPUs. The obtained patterns were meshed and fed into a multi-physics model incorporating mass, momentum, and charge transport processes, developed using the finite element method implemented using Firedrake and PETSc open-source codes to run in HPC environments. Initial results of the developed framework demonstrate the potential of the framework to assess the performance of the NIPS electrodes based on the conditions used to fabricate them.
Study on hydrogen dissociation and adsorption behavior in hydrogen-blended natural gas pipelines
ABSTRACT. Blending hydrogen into natural gas pipelines provides an economic approach for hydrogen transportation on a large scale. Hydrogen embrittlement (HE) poses a great threat to the safety of hydrogen-blended natural gas pipelines. As a precursor to hydrogen cracking, it is crucial to clarify the influencing factors of hydrogen dissociative adsorption. In this study, the dissociative adsorption mechanism of hydrogen in hydrogen-blended natural gas pipelines was investigated by density functional theory (DFT), and the accuracy of the simulation results was verified in conjunction with macroscopic gaseous hydrogen permeation experiments. The hydrogen adsorption energy, dissociation activation energy, electric density difference and partial density of states were calculated to understand the bonding mechanism by modelling hydrogen adsorption on surfaces such as Fe(100) surfaces, grain boundaries, and corrosion product films, and the effect of different surface states on dissociative adsorption was investigated. In addition, the competitive adsorption of O2, CO2, CO and hydrogen was modelled, and the effect of impurity gases in natural gas on dissociative adsorption. The results show that the dissociative adsorption of hydrogen is selective and can only occur at specific adsorption sites, the adsorption energy of hydrogen at grain boundaries is much lower than low-index crystal surfaces, and surface defects are the active regions for hydrogen dissociation. Corrosion product have an inhibitory effect on hydrogen dissociative adsorption and can act as a hydrogen barrier. In addition, CO and O2, which have strong adsorption capacity, will compete with hydrogen for surface adsorption sites, raising the hydrogen dissociation barrier. The results were further validated and analyzed in conjunction with experiments. This study helps to improve the understanding of the mechanism of competitive adsorption of impurity gases on pipeline steel surface, and thus suggests a feasible method to reduce the risk of HE in pipelines by inhibiting hydrogen dissociative adsorption.
Study on consequences and safety distance for leakage and explosion in pipelines of hydrogen-blended natural gas stations
ABSTRACT. Transporting hydrogen-blended natural gas through existing natural gas pipe-line networks is a key strategy for meeting the growing demand for hydrogen energy. However, due to the distinct physical and chemical properties of hydrogen compared to natural gas, hydrogen-blended natural gas stations face unique safety challenges, particularly in terms of pipeline leakage and explosion risks. This study develops a comprehensive model for hydrogen-blended natural gas stations and investigates the effects of various factors, including leak hole size, hydrogen blending ratio, leakage pressure, and ambient wind speed, on the processes of leakage diffusion and explosions. Additionally, the study assesses the applicability of existing safety distance standards in the context of accident consequences. The findings reveal that larger leak hole sizes and higher hydrogen blending ratios significantly accelerate gas diffusion rates and expand the diffusion range, leading to an in-crease in the volume of flammable gas clouds and a corresponding rise in explosion risk. When the hydrogen blending ratio exceeds 50%, the peak overpressure, flame propagation speed, and the impact range of pressure waves during explosions all increase substantially. Higher leakage pressures exacerbate both gas diffusion and explosion intensity, while increased ambient wind speeds reduce gas accumulation within the station but elevate the risk of gas dispersion to surrounding areas. Furthermore, the study highlights that existing safety distance standards, which are based on natural gas leak-age characteristics, are inadequate for hydrogen-blended natural gas stations. While the current safety distance of 15 meters can effectively prevent severe damage to buildings when the hydrogen blending ratio is below 30%, it be-comes insufficient when the ratio exceeds 50%. This research provides a theoretical foundation for the safety design and optimization of standards for hydrogen-blended natural gas stations, offering critical insights for enhancing operational safety and mitigating risks.
A Multilayer and temporal network for studying the connections of cross-listed stocks
ABSTRACT. Multilayer networks can model multiple relationships between different entities and are gradually being applied to the study of financial system connections. In this paper, we employ a multilayer and temporal network framework to quantify the connections among stocks cross-listed in mainland China and Hong Kong, providing a systematic perspective beyond traditional pairwise analyses. Specifically, we construct a multilayer network based on daily closing data from 72 pairs of cross-listed stocks for 2551 trading days, with each layer of the network representing a trading market. Meanwhile, a rolling window approach is used to explore the temporal evolution of these connections, particularly during major global events. In addition, we develop a centrality difference indicator to measure the influence disparity of cross-listed companies in different markets, and perform community detection based on this indicator to mine the characteristics of cross-listed stocks in various communities. Results show that cross-listed stocks exhibit asymmetric influence in different markets, with the core stocks changing over time in both A-share and H-share. The overall connection of multilayer network can be significantly enhanced during systemic financial crises, geopolitical conflicts, and other crisis events. Notably, the community detection results based on centrality difference distributions suggest that the influence disparity of cross-listed companies in different markets may be related to the connection strength and premium levels of their paired stocks. Our study provides a novel framework for a deeply understanding of the dynamic connections among cross-listed stocks, as well as valuable insights for risk management.
A novel routing algorithm for optical networks based on ML methods
ABSTRACT. Routing is essential to the seamless operation of an optical
network. In this paper, we develop a novel routing algorithm specially
tailored for optical networks. It uses machine learning methods to predict the feasibility of a particular route and depending on the prediction
outcome makes the decision on whether to admit or reject the specific
route. The machine learning model for feasibility prediction is created
using data gathered via the control plane from a real telecoms network.
The results obtained show the superiority of the proposed routing algorithm when compared with the standard approach.
Covering the online spectrum of opinion in social context: the benefit of network node sampling through an Italian case study
ABSTRACT. Capturing a diverse range of opinions, sentiments, and topics
is essential when selecting training data for statistical and machine
learning models, particularly those that require interpretability. Online
comments offer a valuable source of public opinion, but they often present
a skewed representation, with opposing viewpoints being overrepresented
compared to supportive ones. This imbalance can lead to biased models
that reinforce stereotypes and reduce fairness and utility. The goal is
to ensure that a broad spectrum of opinions and sentiments is reflected
in the data, helping mitigate bias and providing a more comprehensive
dataset for training. By doing so, we can develop fairer, more transparent
models that are better suited for analysing complex social issues. To
achieve this, it is crucial to employ effective sampling techniques, such
as space-filling sampling on networks, that ensure thorough coverage of
various topics and sentiments in online discussions. We will demonstrate
this methodology with a simulated case study, and analysing social media
comments focusing on online debate around migration. Considering
the limitations of existing Italian lexical resources, we will introduce a
novel sampling technique that ensures both topic and sentiment are adequately
represented in the corpus, enhancing its overall reliability and
breadth.
Decision Trees and Machine Learning for Cybersecurity: How Model Settings Affect Attack Detection
ABSTRACT. The aim of this research is to evaluate the effectiveness of decision trees in detecting cyberattacks and to compare different learning conditions' impact on classification performance. The paper presents an analysis of the impact of various hyperparameters, including splitting criteria (Gini vs. Entropy), feature selection, and tree depth, on the accuracy, precision, recall, and F-measure of models. A comparative analysis is performed using machine learning classifiers, such as Gradient Boosting, AdaBoost, Support Vector Machines (SVM), Lasso, and Random Forest, to assess their relative performance in cyber-attack classification. The findings demonstrate that decision trees are able to achieve high effectiveness in detecting network intrusions, and feature selection can enhance classification performance. Some of the classifiers among the evaluated models, including Random Forest and Gradient Boosting, offer better performance, showcasing their potential as alternatives to decision trees in cybersecurity applications. The results show the importance of hyperparameter optimization and feature engineering, particularly in improving threat detection model performance and/or accuracy.
A machine learning-based framework for predicting candidate drug side effects from biological networks
ABSTRACT. Artificial Intelligence (AI) has emerged as a powerful and effective tool with diverse applications across medicine, biology, pharmacology, and the broader health sciences. An inherent drawback of drug therapies is the potential for side effects, which are adverse reactions that negatively impact human health. These unwanted and undesirable effects typically surface during clinical trials or in clinical practice and require comprehensive investigation. In recent years, AI has been applied in several fields, including pharmacology and pharmacovigilance, for studying and analysing drug side effects, also for the purpose of making predictions of novel candidate adverse effects. Likewise, network science and biological multilayer network analysis has become widely used as an effective and efficient tool for modelling interactions between biological objects.
In this paper, we presented a framework for predicting potential drug side effects using machine learning techniques applied to biological networks. Our key contribution lied in the development of a novel framework that integrated multiple predictive models to infer candidate drug side effects, by combining principles of graph theory with Machine Learning (ML). Experimentation supports the application of the ML-based models implemented in the proposed framework for predicting novel (candidate) drug side effects from biological networks.
Biological Community Detection with Graph neural network and Network Curvature analysis on Gene Co-expression Networks.
ABSTRACT. Biomedical networks are critical for representing complex biological systems, and network curvature is a key structural property that captures topological features not highlighted by traditional graph metrics. This study introduces a Graph Neural Network (GNN)-based approach for detecting communities in cancer-specific Gene Co-expression Networks (GCNs), using \textit{Ollivier-Ricci curvature} as an integral feature. The inclusion of curvature has shown to enhance the detection of biologically significant communities, improve network modularity, and enable finer partitioning. These preliminary results indicate that curvature-based analyses can offer new insights into the organization of gene co-expression networks, aiding in the understanding of biological modularity, disease mechanisms, and functional interactions.
Understanding the Limitations of Deep Transformer Models for Sea Ice Forecasting
ABSTRACT. It would not be an exaggeration to say that we live in the era of transformers. Due to the great results of generative models for video prediction, spatio-temporal data of various kinds are usually treated as video-like sequences - and this is a good assumption for many problems. However, we want to argue that transformer-based prediction is not the best option for some spatio-temporal cases with regular grid and strong periodicity (since most discussions about the limitations of transformer applicability focus only on time series).
In the paper, we considered the task of sea ice forecasting and analyzed two transformer-based architectures (TimeSformer and SwinLSTM) against the proposed baseline - a lightweight convolutional network with different setups of convolutional layers (2D and 3D).
Experiments for long-term forecasting of Arctic seas show that transformers do not reproduce the annual dynamics of sea ice. At the same time, the CNN-based solutions allow to outperform the existing state-of-the-art numerical (SEAS5) and data-driven (IceNet) forecasts, with a quality improvement of up to 30\% in the mean absolute error and up to 10\% in the structural similarity index. A similar experiment is provided for the synthetic example of video data. Due to the analysis of the obtained results, this problem is caused by the nature of the model and the data and can be faced in many scientific and industrial tasks outside sea ice.
Code and supplementary materials for this research are available on GitHub: \url{https://github.com/ITMO-NSS-team/sea_ice_transformers}.
Generation of quality Green’s function libraries in complex three-dimensional crustal structures by adaptive mesh refinement
ABSTRACT. Fault slip estimation from crustal deformation is essential for understanding earthquake mechanisms, and three-dimensional subsurface models have become incorporated in the estimation to account for the geometric and material heterogeneity in the crust. In the estimation process, a precomputed Green’s function library (GFL), which represents the displacement field due to unit point load at each receiver on ground surface, can be employed for computing arbitrary source responses through convolution of the source terms and the GFL. However, challenges remain in generating meshes adapted to the singularity of point loads and in assessing the accuracy of the GFL. We developed a GFL computation method, Adaptive Finite Element Method for Green’s Function Library (AFEM-GFL). By combining initial meshes with good element quality and a mesh refinement algorithm that is resistant to element quality degradation, our method can generate meshes well adapted to computing GFL. The accuracy of the results is assessed through convergence of the solution and comparisons between the convergent solutions from different initial meshes. In numerical experiments on a two-layered half-space and a realistic crustal structure of Japan, accurate and convergent GFLs were obtained with moderate amount of computational resources in the settings where it was difficult to achieve with uniform meshes even using a massively parallel supercomputer. These quality GFLs will serve as a robust foundation for reliable fault slip estimation.
Large-scale nonlinear viscoelastic simulation for crustal deformation accelerated by data-driven method and multi-grid solver
ABSTRACT. We developed a fast method for highly detailed nonlinear viscoelastic crustal deformation analysis by a combination of a data-driven method and a multi-grid solver. Here, highly accurate estimations of the solution of the next time step are obtained using a data-driven predictor based on the history of past solutions, which reduces the number of iterative solver iterations and thus reduces the computation cost. Although this method has been shown to enable fast linear viscoelastic analysis, its validity has not been confirmed for nonlinear viscoelastic problems. Numerical experiments have shown that the data-driven method reduces the number of iterations by 3.35-fold. To achieve further acceleration, we introduced a multi-grid solver capable of efficiently solving large systems of equations. The proposed method with the data-driven method and multi-grid solver is applied to nonlinear viscoelastic crustal deformation analysis of the Nankai Trough region, and it is shown that the proposed method achieved a 15.1-fold speedup, which enabled many large-scale crustal deformation simulations within reasonable computational costs. The fast nonlinear viscoelastic analysis of highly detailed crustal struc- ture models enabled by this study is expected to contribute to the advance of interplate state estimation.
Simple error estimation for PIES in 2D elasticity problems
ABSTRACT. The paper presents a posteriori error estimation strategy in the parametric integral equation system (PIES). It uses collocation point differences between solutions obtained numerically by PIES and another solutions interpolated based on the ini-tial PIES analysis. Various techniques for interpolation are proposed: by repeat-ing the interpolation omitting one collocation point, by interpolating using only values from adjacent collocation points and by the one degree higher polynomial obtained using the least squares approximation. This allows the calculation of lo-cal and global percentage error using the integral over the mentioned differences. Finally, it can be applied to ensure convergence of the solutions using the PIES method or to adaptive refinement of distribution or number of collocation points by identifying boundary regions where the error is relatively high.
Using B-spline function properties in the PIES method to handle singularities in boundary value problems
ABSTRACT. The paper presents the formulation of a parametric integral equation system (PIES) for 2D boundary value problems with singular solutions. The proposed approach combines B-spline basis functions with the PIES formalism to approximate the results in challenging problems such as notched plates or concentrated loads. B-spline function with a specific degree and knots is used in approximating series instead of previously applied polynomials (such as Lagrange or Chebyshev). Increasing knot multiplicity to reduce continuity at some locations or increasing the number of knots in singular regions allows for more accurate results. The proposed approach is validated through selected problems. The results confirm that the B-spline-enhanced PIES approach can effectively solve problems with singularities with improved accuracy.
Comparison of crash simulations on two types of flying cars
ABSTRACT. Abstract: Numerical simulations are being used to ensure safety in the development of flying cars, which are expected to solve all kinds of traffic problems. In order to reproduce flight in a virtual environment, this research has been conducting numerical simulation research using the moving computational domain method and the multi-axis sliding mesh method to calculate the coupling of fluid and rigid body motion. This method enables rotational motion of the rotor and flight in an infinite region, and visualization of the flow field around the aircraft and its behavior including control. In this study, a comparison of the behavior during sudden rotor stops was performed on two different aircraft: the coaxial contra-rotating octorotor eVTOL treated in a previous study and a newly modeled domed dodecarotor eVTOL. The increased number of rotors and the wider circular arrangement of the rotors allowed for a wider range of measures to be taken in the event of rotor stoppage, and the crash risk was reduced by a factor of approximately 1/100. The differences in crash conditions and changes in aircraft attitude based on a comparison of the two models indicate that this calculation method allows for design improvements and behavior prediction based on relative evaluation.
Domain solutions obtained by the FPIES for potential 2D BVPs
ABSTRACT. The paper presents the fast parametric integral equations system (FPIES) for domain solutions of potential 2D boundary value problems (BVPs). The FPIES has been successfully applied in modelling 2D potential BVPs and finding solutions on the boundary. The combination of the modified fast multipole technique with the PIES reduced the numerical computation time and RAM usage in the fast PIES. Similar techniques are used to find solutions in the domain. The method is demonstrated with the solution in the domain of two BVPs.
Physics-Aware Compression of Plasma Distribution Functions with GPU-Accelerated Gaussian Mixture Models
ABSTRACT. Data compression is a critical technology for large-scale plasma simulations. Storing full particle information requires Terabyte-scale data storage and analysis requires ad-hoc scalable post-processing tools. We propose a physics-aware in-situ compression method using Gaussian Mixture Models (GMMs) to approximate electron and ion velocity distribution functions with a number of Gaussian components. This GMM-based method allows us to capture plasma features such as mean velocity, temperature, and the possibility of identifying heating processes and generation of beams. To reduce computational overhead, we first construct a histogram and apply GPU-accelerated, in-situ GMM fitting within iPIC3D, a large scale implicit Particle-in-Cell simulator, ensuring real-time compression. The compressed representation is stored using the ADIOS 2 library, thus reducing I/O costs. The GPU and histogramming implementation provides a major speed-up with respect to GMM on particles (both in time and in memory), enabling real-time compression. Compared to algorithms like SZ, MGARD and BLOSC2, our GMM-based method has a physics-based approach, retaining physical interpretation plasma phenomena such as beam formation, acceleration and heating mechanisms while reducing storage and I/O overhead.
Meta-Instance Selection. Combining the Properties of Multiple Instance-Selection Methods.
ABSTRACT. This study presents two novel approaches for developing a multiple-meta-instance selection method, an advanced algorithm designed for efficient pruning of training samples in classification problems. The proposed meta-instance selection framework reformulates the traditional instance selection problem by introducing a meta-feature space, a problem-agnostic representation space. The transformation enables instance selection to be framed as a classification task in the meta-feature space, facilitating efficient computation with a time complexity of O(n log(n)). A standard classification algorithm, such as Random Forest, can then be employed in the meta-feature space to determine the inclusion or exclusion of individual samples.
To enhance performance, we explore two strategies for combining multiple meta-instance selection algorithms: (1) constructing an ensemble of meta-classifiers and (2) concatenating many meta-sets. Experimental evaluations demonstrate that the meta-set concatenation approach surpasses both classical instance selection techniques and existing meta-instance selection methods. Moreover, the proposed algorithm significantly accelerates the instance selection process—achieving even by two or three orders of magnitude speed-up, depending on dataset size and the reference instance selection method.
Preliminary comparison of different EDs performance, using Simulation
ABSTRACT. A reference model is a model of something that contains a fundamental objective or idea of something and can be established as a reference for multiple purposes. We performed an analysis based on a reference model following the Spanish Ministry of Health's document of standards and recommendations to achieve better care, efficiency, and uniformity in emergency services. This standard describes the guidelines and resources needed in hospital emergency services for better patient care. Adopting this standard should facilitate better emergency care for patients of all ages once analyzed. We also analyzed the standards of emergency services in the following countries: the United States, England, Germany, Canada, Paraguay, Argentina, and the United Arab Emirates. The objective of the research is to analyze the efficiency of the Spanish model following its reference standard compared to the standards of other countries, to explore how the Spanish emergency service would work using the parameters of the emergency service standards compared to the standards of different countries, specifically the KPI we analyzed is the door to the doctor (DtD), through simulation. It was concluded that in some countries, DtD times improve compared to the Spanish reference standard, and in some cases, they worsen.
Improving Object Detection Quality in Football Through Super-Resolution Techniques
ABSTRACT. This research examines the effectiveness of super-resolution techniques in improving object detection accuracy in football. Given the sport's fast pace and the need for precise tracking of players and the ball, super-resolution can offer significant improvements. The study applies super-resolution techniques to SoccerNet football videos and evaluates their impact on Faster R-CNN detection accuracy. Findings reveal a significant boost in object detection accuracy following the application of super-resolution preprocessing. Enhancing object detection by integrating super-resolution techniques provides substantial advantages, particularly in low-resolution settings, with a 12\% rise in mean Average Precision (mAP) at an IoU (Intersection over Union) range of 0.50:0.95 for 320x240 pixel images when the resolution is quadrupled using RLFN. As the image dimensions grow, the extent of improvement becomes less pronounced; however, a consistent enhancement in detection quality remains clear.
Moreover, the implications of these results for real-time sports analytics, player tracking, and the overall viewing experience are discussed.
To select or not to select? The role of meta-features selection in meta-learning tasks with tabular data
ABSTRACT. In meta-learning tasks with tabular data, the choice of meta-features significantly impacts model performance and interpretability. This study investigates the necessity and methods of meta-feature selection in the context of meta-learning, particularly for tabular datasets. We address the fundamental question: Is it better to select a subset of meta-features or use the entire feature set? We examine various selection techniques, including filter, wrapper, and embedded methods, as well as a novel causal-based approach utilizing counterfactual reasoning. Our experiments demonstrate that feature selection generally enhances performance, with causal-based methods, especially those leveraging counterfactual generation, showing superior efficiency and generalizability. Furthermore, we explore how these methods fare under shifts in data, particularly when non-informative features are added. The results reveal that the counterfactual method maintains high efficacy across different meta-learners and exhibits a favorable balance between model performance and interpretability. These findings underscore the importance of meta-feature selection in improving the adaptability and transparency of meta-learners for tabular data tasks.
Code and supplementary materials for this research are available on GitHub: https://github.com/ITMO-NSS-team/MetaSelect.
Backtranslation and paraphrasing in the LLM era? Comparing data augmentation methods for emotion classification.
ABSTRACT. Numerous domain-specific machine learning tasks struggle with data scarcity and class imbalance. This paper systematically explores data augmentation methods for NLP, particularly through large language models like GPT. The purpose of this paper is to examine and evaluate whether traditional methods such as paraphrasing and backtranslation can leverage a new generation of models to achieve comparable performance to purely generative methods. Methods aimed at solving the problem of data scarcity and utilizing ChatGPT were chosen, as well as an exemplary dataset. We conducted a series of experiments comparing four different approaches to data augmentation in multiple experimental setups. We then evaluated the results both in terms of the quality of generated data and its impact on classification performance. The key findings indicate that backtranslation and paraphrasing can yield comparable or even better results than zero and a few-shot generation of examples. The best results were achieved with backtranslation using the DeepL model. Future research should explore experiments with different models and assess the effectiveness of data augmentation across various datasets and tasks. This approach promises broader applicability.
A test space refinement strategy for Robust Variational Physics-Informed Neural Networks
ABSTRACT. In early 2024, the Robust Variational Physics-Informed Neural Networks (RVPINNs) method was introduced in [1] as a robust extension of the Variational Physics-Informed Neural Networks (VPINNs) method [2]. Unlike standard VPINNs, RVPINNs minimize a loss based on the discrete dual norm of the residual, providing a reliable estimator of the approximation error in the energy norm under the assumption of a local Fortin operator. This leads to improved accuracy in approximating the partial differential equations governing experimental data. However, a key challenge of RVPINNs is the need to invert a Gram matrix at each nonlinear solver step, which can make the method computationally expensive if the variational formulation and discrete test space are not carefully chosen.
In this talk, we will present an adaptive strategy for efficiently defining the discrete test space when using standard finite element functions. The approach begins with a coarse mesh for the test space, followed by an iterative process where the nonlinear solver is applied until a predefined oscillation term is reached. The test space mesh is then refined using an a posteriori error estimator, and this process continues until convergence is achieved. We will show how this adaptive strategy improves the accuracy of the neural network approximation, while significantly reducing computational cost when compared to considering a fine uniform mesh for the discrete test space at the beginning of the training. These advantages will be illustrated through a series of numerical experiments on challenging diffusion-advection-reaction problems.
References
1. Rojas, S., Maczuga, P., Muñoz-Matute, J., Pardo D., Paszyński, M.: Robust Variational Physics-Informed Neural Networks. Computer Methods in Applied Mechanics and Engineering 425, 116904 (2024)
2. Kharazmi E., Zhang Z., Karniadakis GE.: Variational physics-informed neural networks for solving partial differential equations. arXiv preprint arXiv:1912.00873 (2019)
Augmenting Petrov-Galerkin method with optimal test functions by DNN learning the inverse of the Gram matrix
ABSTRACT. The Petrov-Galerkin (PG) method is a robust alternative to the Galerkin method for finite element simulations of challenging partial differential equations (PDEs).
The solution of the Galerkin method is obtained from the linear system Bx=F resulting from discretization of trial and test spaces. While the Galerkin method enforces the equality of trial and test spaces U_h=V_h and relies on the inf-sup stability condition to ensure solution accuracy, it often fails for difficult problems where the discrete inf-sup constant alpha_h significantly deviates from the abstract inf-sup constant alpha.
This discrepancy leads to numerical instability and incorrect solutions. The PG method addresses this by allowing distinct trial and test spaces (U_h != V_h), enabling the selection of test functions that improve the discrete inf-sup constant alpha_h.
Of particular interest is the Petrov-Galerkin method with optimal test functions (PGO), where the test functions are computed to maximize alpha_h, ensuring stable solutions even for ill-conditioned problems. The PGO method modifies the discrete test space to approximate the abstract stability properties as closely as possible. The computation of optimal test functions involves solving GW=B, where B is the Galerkin matrix, and G is the Gram matrix of the test space's inner product. Solving
B^TWx=W^TF then yields a stable solution. However, the added computational cost of inverting G^-1 poses a significant overhead.
In this work, we propose a novel approach that leverages deep neural networks (DNNs) to approximate the inverse of the Gram matrix for a class of advection-diffusion problems with variable diffusion coefficients. By training the DNN to predict G^-1, we eliminate the computational overhead of matrix inversion, enabling efficient and stable solutions of PDEs. Our results demonstrate the effectiveness of the DNN-enhanced PGO method in maintaining stability and accuracy, even for difficult computational problems where the standard Galerkin method fails. This approach represents a significant advancement in the practical applicability of the Petrov-Galerkin framework for solving complex PDEs.
Graph grammar model for h-adaptation for meshes with quadrilateral, pentagon, and hexagon elements
ABSTRACT. The paper introduces a hypergraph grammar for modeling
mesh adaptation in 2D meshes with quadrilateral, pentagon, and hexag-
onal elements. In this approach, the finite element mesh is represented
as a hypergraph, with all mesh transformations defined by hypergraph
grammar rules. These rules enable the execution of the h-adaptation of
elements, namely the mesh refinement.
EXPBrain: Exponential Integrators for Glioblastoma Brain Tumor Simulations
ABSTRACT. In this paper we discuss a MATLAB implementation of the exponential integrators method employed for simulating of the brain tumor progression. As the input data we utilize publicly available T1 weighted magnetic resonance imaging dataset ds003826, representing healthy individuals. The data from these datasets are originally stored using NIfTI format. We select randomly one anonimized individual from the considered dataset. We normalize the brain scan data using min max normalization to a range of 0 to 255. In the data from the dataset ds003826 the voxel resolution is not isotropic in all directions, so we interpolate the data from dimensions 176×248×256 into 194×248×256 in order to have proper proportions of the human brain. We set the data as a sequence of 256 PNG les with the resolution of 194×248. Having the MRI scan data, we run the exponential integrators method simulating the glioblastoma tumor growth using the Fisher-Kolmogorov diffusion-reaction model with logistic growth. We assume the initial tumor location and run the simulation predicting two years forward tumor growth. For the spatial discretization we employ the nite di erence method, and for the temporal discretization we use the ultra-fast exponential integrators method. Our simulator generates the simulational results suitable for visualization using the ParaView tool.
Introducing B-spline basis functions in neural network approximations
ABSTRACT. In the finite element method (FEM), the solutions of Partial Differential Equations (PDEs) are approximated using linear combinations of prescribed basis functions. The coefficients of the linear combinations are obtained by solving a system of linear equations. The FEM allows for the solution of a PDE for fixed values of the PDE parameters. It is not possible to obtain "at once" the family of solutions of parametric PDEs using FEM. We proposed to introduce B-spline basis functions into neural network approximations, where the coefficients of the basis functions used to approximate the solution are predicted by a neural network. The coefficients are obtained by minimizing the loss functions being the residual of the parametric PDE. Direct approximation of B-spline coefficients by NN has several advantages compared to standard FEM. First, it allows us to obtain a family of solutions of the parametric PDE "at once". The PDE parameters are input to the neural network, and the output involves the coefficients of the basis functions.
Second, it allows obtaining the solution of a parametric PDE without the construction and solution of a system of linear equations. Third, since neural networks are universal approximators, direct approximation of B-spline coefficients by NN may find a dependence between the PDE parameters and the coefficients of the basis functions used to approximate the solution. We compare our method of direct approximation of B-spline coefficients by NN with the Physics-informed Neural Networks (PINN). The training of PINN requires "stretching" of the non-linear function, the neural network, into the multi-dimensional manifold of the parametric PDE solutions. Our method uses the linear combination of fixed basis functions to approximate the solution. The training of our method requires learning the dependence between the PDE parameters and the basis functions' coefficients. The approximations of B-spline coefficients by NN inherit all the features of standard FEM approximations. Our method is discussed with all the details and comparisons to PINN and FEM solvers.
A Computational Framework for Modelling Biomechanical Tumour Dynamics and Tissue Interactions: A Proof-of-Concept in Pleural Mesothelioma
ABSTRACT. In Malignant Pleural Mesothelioma (MPM), solid stress and tissue deformation significantly impact tumour growth and invasion. This study presents a computational framework that integrates biomechanical tumour dynamics, tissue deformation, and force interactions within a realistic anatomical setting. Using the Finite Element Method, the framework is applied to a lung mesh reconstructed from CT scans, incorporating a synthetic mesothelioma tumour with defined material properties. Numerical results from the simulations closely match the analytical solution, with deviations within 5%, confirming the model’s reliability and accuracy. Simulations of point compression and surface expansion effectively capture the localised tumour deformation and lung volume changes, replicating expected breathing mechanics under different conditions. The findings emphasize the role of mechanical interactions in tumour progression, demonstrating how increased tissue stiffness affects deformation patterns and respiratory dynamics. This study establishes a foundation for integrating computational biomechanics with predictive tumour modelling, offering potential applications in personalised medicine for MPM.
Towards sensitivity analysis: 3D venous modelling in the lower limb
ABSTRACT. Deep vein thrombosis (DVT) of the lower extremity fre-
quently leads to long-term complications known as post-thrombotic syn-
drome (PTS). The current clinical workflow for DVT and PTS treat-
ment lacks sufficient evidence. The significance of the variation in the
venous anatomy is yet to be understood. We report an analysis of a
set of idealised 3D geometries of iliac vein unification to assess the im-
portance of shape variability, inflow conditions, and viscosity on local
haemodynamics - specifically on the wall shear stress metrics. Regions of
low average wall shear stress and high oscillating shear index have been
associated with prothrombotic effects on the walls of blood vessels. A
detailed steady state analysis focused on the low wall shear stress distri-
butions below three thresholds (< 0.15P a, < 0.10P a < 0.05[P a]). The
preliminary work in the transient state focused on the oscillating shear
index above three thresholds (> 0.25, > 0.35, > 0.45). We found that all
the variations implemented had an effect on the size and shape of the
absolute vein wall area subject to the shear metrics of choice under the
assumed flow conditions. The results obtained in this research will serve
as a basis for the interpretation of patient-specific geometries of the iliac
vein unification affected by deep vein thrombosis.
Accelerating Two-Dimensional k-Wave Ultrasound Simulations Through Pruned FFT: A Treatment Planning Optimisation
ABSTRACT. Wave propagation simulations are foundational tools across scientific and medical applications, yet their computational demands become significant for high-resolution simulations, particularly in medical applications where precise representation of different tissue geometries is crucial. This paper presents a novel approach to accelerate 2D wave propagation simulations in the k-Wave toolbox. Our method focuses on optimising Fourier transform computations through spectrum pruning. The Acoustic Field Propagator along with a bisection algorithm to estimate the position of the spectral coefficients is used. Through these optimisations, our approach achieves significant performance gains, demonstrating speedups of up to 1.8x for large simulation domains. Experimental evaluation on medical ultrasound simulations demonstrates that the proposed method achieves focal point errors below 1% with minimal focus position shifts, while skipping up to 90% of spectral coefficients in large domains. This results in a significant simulation time reduction by half over the large simulation domains. Although the proposed method primarily focuses on accelerating k-Wave toolbox wave propagation simulation, it could be generally applied to wave propagation problems.
Predicting disease transmission rates for hybrid modeling of epidemic outbreaks: statistical and machine learning approaches
ABSTRACT. Hybrid disease modeling is a perspective area of research that allows using detailed individual-based models for the outbreak onset phase and lightweight compartmental models to capture the general trend of the disease progression. In such a way, the method of hybrid modeling provides a good trade-off between the simulation speed and the accuracy of reproducing disease dynamics. One of the problems related to this approach is how to switch properly between the two models. That included detecting the right time moment to finish simulations with the detailed model and calculating correctly the input parameters for the compartmental model. In this paper, we propose an implementation of switching which relies on evaluation and prediction of disease transmission rate. Using an example with a network-based model and a discrete compartmental model, we demonstrate several methods of disease transmission prediction based on statistical models and machine learning approaches, and analyze their advantages and disadvantages. The developed methods can be generalized to hybrid modeling of highly detailed demographic processes and propagation processes in general.
Lightweight heterogeneous SEIR models for epidemic surveillance in Russian cities: turning synthetic populations into equations
ABSTRACT. Influenza and other acute respiratory diseases pose a significant challenge to global health. The complexity of analyzing and mitigating influenza transmission is related to heterogeneity of contact network structures in modern cities. The need for effective public health strategies has driven the development of highly detailed network and agent-based models. To overcome a drawback of modeling multi-agent systems, which is their high demand for computational resources, approximate models can be employed. In our paper, we present an approach that allows to convert heterogeneous synthetic populations into an input for the edge-based compartmental SEIR model. We demonstrate the method application by simulating influenza spread in a contact network of the synthetic population of St. Petersburg, Russia. At a cost of neglecting some details in contact network structure, the proposed algorithm allows to greatly enhance simulation speed compared to multi-agent modeling, and at the same time to preserve population heterogeneity, which makes it a good choice for application in epidemic surveillance.
{\tt DarcyLite} Modules for Property-preserving Transport Solvers
ABSTRACT. For solute transport in porous media
modelled by the time-dependent convection-diffusion equation,
positivity of concentration is
an important property a numerical solver should respect.
In this paper,
we investigate a new finite volume solver on quadrilateral meshes.
The solver uses mapped $ Q_1 $ bilinear polynomials
for spatial approximation of concentration.
Upwinding and flux correction are also used
in combination with the implicit Euler for time-marching
to maintain nonnegativity of numerical concentration.
The solver is robust in the sense that
it can handle convectional dominance very well.
Code modules based on efficient implementation of the solver in \texttt{Matlab}
are incorporated in our package \texttt{DarcyLite}.
Numerical experiments are presented to illustrate the performance of this new solver.
Finite volume method
An Iterative Scheme for the Solidification and Macro-segregation Benchmark Modeling
ABSTRACT. The processes of solidification and macro-segregation involve intricate interactions across multiple physical, phase, and compositional fields, including mass, momentum, energy, and material transfer. Accurate prediction of phase transitions, chemical heterogeneities, and compositional flows is crucial in fields such as materials science, energy science, and planetary science. Numerical benchmark studies provide an effective means to explore these phenomena. This paper presents an iterative scheme based on operator splitting and evaluates its accuracy, stability, and implementation through a relevant benchmark problem. The results demonstrate strong performance of the scheme, particularly in capturing key physical phenomena such as channel segregation, freckle formation, and edge effects.
Natural convection in periodically heated porous-fluid system under local thermal non-equilibrium conditions: a numerical study for enhanced thermal management
ABSTRACT. This numerical study investigates natural convection and heat transfer in a closed chamber with porous medium. The system combines a porous layer and a Newtonian fluid with temperature-dependent viscosity, subjected to time-dependent thermal excitation. Governing equations, formulated in dimensionless stream function and vorticity variables, integrate mass, momentum, and energy conservation using the Darcy-Brinkman model and Boussinesq approximation. The LTNE framework resolves thermal decoupling between the porous matrix and fluid, overcoming limitations of local thermal equilibrium assumptions. A finite difference numerical scheme is employed to solve the dimensionless equations, analyzing the interplay of LTNE parameters (interphase heat transfer, parameters of solid structure) and periodic heating (frequency, amplitude). Results demonstrate that LTNE conditions significantly alter thermal stratification, velocity asymmetry, and heat transfer rates (quantified with help Nusselt number). Elevated heating frequencies suppress convective instabilities, while variable viscosity amplifies thermal gradients. The porous-fluid conductivity ratio critically modulates thermal non-equilibrium, with lower ratios exacerbating temperature disparities. This work validates the necessity of LTNE models for systems involving rapid thermal transients, heterogeneous media, or variable properties. The findings provide critical insights for optimizing thermal management in energy storage, electronic cooling, and geothermal systems.
A New Technique for Enhanced Monochrome Visualization of Non-Visual Data
ABSTRACT. This paper introduces a novel, unconventional method to enhance the visual clarity of monochrome images derived from non-visual data sources. The approach involves a two-step process: image pseudo-colorization followed by decolorization. Surprisingly, this counterintuitive technique can significantly improve discernibility of image features, irrespective of their size, shape, or original visual prominence. The paper delves into the algorithmic details of this method and presents experimental results on a representative dataset of IR, X-ray, MRI, and ultrasound images. When disregarding factors related to natural image appearance (which are irrelevant in non-visual domains), this method outperforms conventional image enhancement techniques, including sophisticated ones, in terms of standard image quality criteria, i.e., sharpness, contrast, and overall detail perceptibility. This superiority is substantiated by both subjective evaluations and objective metrics. The success of this technique hinges on the careful selection of color maps and the application of a specific, recently proposed decolorization scheme. The technique is well-suited for various visual data analysis tasks in non-visual domains (where the concept of image naturalness is less pertinent), including AI-based solutions.
SupResDiffGAN a new approach for the Super-Resolution task
ABSTRACT. In this work, we present SupResDiffGAN, a novel hybrid architecture that combines the strengths of Generative Adversarial Networks (GANs) and diffusion models for super-resolution tasks. By leveraging latent space representations and reducing the number of diffusion steps, SupResDiffGAN achieves significantly faster inference times than other diffusion-based super-resolution models while maintaining competitive perceptual quality. To prevent discriminator overfitting, we propose adaptive noise corruption, ensuring a stable balance between the generator and the discriminator during training. Extensive experiments on benchmark datasets show that our approach outperforms traditional diffusion models such as SR3 and I$^2$SB in efficiency and image quality. This work bridges the performance gap between diffusion- and GAN-based methods, laying the foundation for real-time applications of diffusion models in high-resolution image generation.
Enhancing AI Face Realism: Cost-Efficient Quality Improvement in Distilled Diffusion Models with a Fully Synthetic Dataset
ABSTRACT. This study presents a novel approach to enhance the costto-quality ratio of image generation with diffusion models. Our solution
introduces a fully synthetic pairwise dataset of images from distilled and
undistilled versions of a model (i.e. FLUX.1-schnell and FLUX.1-dev).
We hypothesize that differences between distilled and undistilled models
are consistent and, therefore, learnable within a specialized domain, like
portrait generation. Then, we train an image-to-image translation head.
The proposed method works without requiring manual annotations or
real reference photos. Using two sets of low- and high-quality synthetic
images, our model is trained to refine the output of a baseline generator
(e.g., FLUX.1-schnell) to a level comparable to state-of-the-art models
like FLUX.1-dev, which are more computationally intensive.
We train and compare multiple variations of our image-to-image model,
both pairwise (U-Net) and non-pairwise (CycleGAN, ESA-CycleGAN).
Our results show that the pipeline, which combines a distilled version
of a large generative model with our enhancement layer, delivers similar
photorealistic portraits to the non-distilled version with up to an 82%
decrease in computational cost compared to FLUX.1-dev. This study
demonstrates the potential for improving the efficiency of AI solutions
involving large-scale image generation.
Transferability of UNet-Based Downscaling Model for High-Resolution Temperature Data Across Diverse Regions
ABSTRACT. High-resolution atmospheric data is essential for understanding local atmospheric processes, however it is computationally expensive to achieve such high resolutions through physical models. Recently, deep learning techniques, particularly those used in Single Image Super-Resolution, have emerged as a promising approach for statistical downscaling. However, much of the existing research has focused on enhancing model performance within small geographical regions, with limited attention given to the transferability of these models to diverse areas outside of their training domain. This paper introduces a methodology that evaluates the ability of a UNet model to downscale daily 2-meter temperature data outside its training region. The proposed approach uses one-third of the Contiguous United States to train the model, and assesses its performance on unseen areas. Our experimental design deliberately tests both spatial and temporal generalization, demonstrating that relatively compact models can effectively transfer downscaling capabilities to new regions. This results in improvements across key performance metrics including Mean Absolute Error, Root Mean Square Error, and Peak Signal-to-Noise Ratio. Additionally, our approach significantly reduces computational costs while improving downscaling accuracy across diverse climatic and topographic conditions.
Bat Algorithm for Automatic Chaos Control Method Driven by Multiplicative Pulses to the System Variables on the Logistic Map
ABSTRACT. Chaotic systems are characterized by extreme sensitivity to initial conditions, where two arbitrarily close starting points lead to exponentially divergent trajectories over time. Since the 1990s, research has demonstrated that chaotic behavior can be controlled, a phenomenon known as chaos control. In this paper, we address the problem of chaos control in unidimensional maps, specifically focusing on stabilizing chaotic dynamics to a periodic orbit of a given period. Our approach builds on a previously proposed method that applies control pulses of intensity $\lambda$ to the system variables every $\Delta n$ iterations, where $\lambda$ and $\Delta n$ are adjustable parameters. We formulate this problem as a challenging multimodal, multivariate, continuous, nonlinear optimization task and tackle it using the bat algorithm, a popular swarm intelligence method. To evaluate the effectiveness of our approach, we conduct computational experiments on the logistic map under various parameter settings. The results indicate that our method performs effectively for all tested chaotic behaviors. We conclude that the proposed approach is a promising step toward an automated procedure for chaos control in chaotic maps.
Multiscale Parallel Simulation of Malignant Pleural Mesothelioma via Adaptive Domain Partitioning – an Efficiency Analysis Study
ABSTRACT. A novel parallel efficiency analysis on a framework for simulating the growth of Malignant Pleural Mesothelioma (MPM) tumours is presented. Proliferation of MPM tumours in the pleural space is simulated using a Cellular Potts Model (CPM) coupled with partial differential equations (PDEs). Using segmented lung data from CT scans, an environment is set up with artificial tumour data in the pleural space, representing the simulation domain, onto which a dynamic bounding box is applied to restrict computations to the region of interest, dramatically reducing memory and CPU overhead. This adaptive partitioning of the domain enables efficient use of computational resources by reducing the three-dimensional domain over which the PDEs are to be solved. The PDEs, representing oxygen, nutrients and cytokines, are solved using the finite volume method with a first-order implicit Euler scheme. Parallelization is realized using the public Python library mpi4py in combination with LinearGMRESSolver and PETSc for efficient convergence. Performance analyses have shown that parallelization achieves a reduced solving time compared to serial computation. Also, optimizations enable efficient use of the available memory and improved load balancing amongst the processors.
Novel Hierarchical Decision Tree Frameworks Introducing Tree Method Bagging-Stump Integration and Height Optimization
ABSTRACT. The paper introduces a novel hierarchical decision tree framework designed to enhance classification quality in dispersed and fragmented data. By integrating two levels of modeling, the framework employs local decision trees to generate prediction vectors, which are then synthesized through a global decision tree for final classification. Five distinct approaches -- Tree, Tree \& height, Bagging \& stump, Bagging \& height, and Bagging -- are proposed and evaluated. Each method varies in how local models are constructed, focusing on factors such as tree depth, bagging methods, and tree stumps. Experimental results on data sets from the UCI Machine Learning Repository demonstrate that the Bagging approach, particularly with an optimized number of bags and trees height, consistently achieves superior performance across metrics including accuracy, F-measure, and balanced accuracy. These findings highlight the framework's robustness and effectiveness in managing dispersed data, offering significant potential for applications in high-dimensional, fragmented and multi-class classification scenarios.
Node-level Performance of Adaptive Resolution in ls1 mardyn
ABSTRACT. In this work we present a node-level performance analysis of an adaptive resolution scheme (AdResS) implemented in ls1 mardyn. An introduction to AdResS is given, together with an explanation of the coarsening technique used to obtain an effective potential for the coarse molecular model, i.e., the Iterative Boltzmann Inversion (IBI). This is accompanied by details of the implementation in our software package,
as well as an algorithmic description of the IBI method and the simulation workflow used to generate results, which might be interesting for practitioners. Results are provided for a pure Lennard-Jones tetrahedral molecule coarsened to a single site, validated by verifying the correct reproduction of structural correlation functions, e.g. the radial distribution function. The performance analysis is based on a literature-based methodology, which provides a theoretical estimate for the speedup based on a reference simulation and the size of the full particle region. Additionally, a strong scaling study was performed at node level. In this sense,
several configurations with vertical interfaces between the resolution regions are tested, where different resolution widths are benchmarked. A comparison between several linked cell traversal routines, which are provided in ls1 mardyn, was performed to showcase the effect of algorithmic aspects on the adaptive resolution simulation and on the estimated performance.
Static Load Balancing for Molecular-Continuum Flow Simulations with Heterogeneous Particle Systems and on Heterogeneous Hardware
ABSTRACT. Load balancing in particle simulations is a well-researched field, but its effect on molecular-continuum coupled simulations is comparatively less explored.
In this work, we implement static load balancing into the macro-micro-coupling tool (MaMiCo), a software for molecular-continuum coupling, and demonstrate its effectiveness in two classes of experiments by coupling with the particle simulation software ls1 mardyn. The first class comprises a liquid-vapour multiphase scenario, modelling evaporation of a liquid into vacuum and requiring load balancing due to heterogeneous particle distributions in space.
The second class considers execution of molecular-continuum simulations on heterogeneous hardware, running at very different efficiencies. After a series of experiments with balanced and unbalanced setups, we find that, with our balanced configurations, we achieve speedups of 44\% and 55\% respectively.
FUMEplot: a Prototype Tool for Automated Visualisation of Uncertainties in Ensemble Modelling Outputs
ABSTRACT. Here we present FUMEplot (Facilitated visualisation of Uncertainty in Model Ensemble outputs using plots), a general-purpose library which facilitates the automated visualisation of ensemble modelling outputs. We demonstrate the tool by applying it to three existing agent-based modelling tools, showcasing its ease of (automated) use and potential for wider application. We also describe how visualisations with FUMEplot can be generated automatically with minimal user configuration effort through its integration with the FabSim3 automation framework.