ABSTRACT. Epidemiological analysis and mathematical modelling have played an influential role in informing the global response to the COVID-19 pandemic. This lecture will provide an introduction to basic concepts and approaches, distinguishing between inferential modelling used to understand past trends and forward-looking (often counterfactual) modelling used to explore different epidemiological and policy scenarios. With that lens, I will review the work leading to the early characterisation of the threat posed by COVID-19 and which informed response measures world-wide. I will then discuss the difficulties posed for countries in sustaining control later in 2020 and how modelling helped to characterise new viral variants and to map the gradual relaxation of social distancing as vaccination was rolled out in 2021. Next, I will discuss the benefits of using multiple models with differing levels of complexity to address critical questions and the technical gaps in simulations of population dynamics revealed by the pandemic. I will conclude with thoughts on research priorities and lessons learned for the application of modelling and simulation in future pandemics and to other global threats.

ABSTRACT. Modeling of turbulent combustion system requires modeling the underlying chemistry and the turbulent flow. Solving both systems simultaneously is computationally prohibitive. Instead, given the difference in scales at which the two sub-systems evolve, the two sub-systems are typically (re)solved separately. Popular approaches such as the Flamelet Generated Manifolds (FGM) use a two-step strategy where the governing reaction kinetics are pre-computed and mapped to a low-dimensional manifold, characterized by a few reaction progress variables (model reduction) and the manifold is then “looked-up” during the runtime to estimate the high-dimensional system state by the flow system. While existing works have focused on these two steps independently, we show that joint learning of the progress variables and the look-up model, can yield more accurate results. We propose a deep neural network architecture, called ChemTab, customized for the joint learning task and experimentally demonstrate its superiority over existing state-of-the-art methods.

ABSTRACT. The paper presents a new modification of the fast parametric integral equations system (PIES) by application of interval numbers and interval arithmetic in solving potential 2D boundary value problems with complex shapes. Obtained interval modified fast PIES is used to model the uncertainty of measurement data, which are necessary to define boundary shape. The uncertainty was defined using interval numbers and modelled using modified directed interval arithmetic previously developed by the authors. The reliability and efficiency of the interval modified fast PIES solutions obtained using such arithmetic was verified on 2D complex potential problems with polygonal domains. The solutions were compared with the interval solutions obtained by the interval PIES. All performed tests indicated high efficiency of the interval modified fast PIES method.

ABSTRACT. Finding a path in a narrow passage is a bottleneck for randomised sampling-based motion planning methods. This paper introduces a technique that solves this problem. The main inspiration was the method of exit areas for cavities in protein models, but the proposed solution can also be used in another context. The proposed method detects whether a collision-free path through a narrow passage exists at all. With such information, it is possible to quit the motion planning computation if no solution exists and its further search would be a loss of time. The method was tested on real biomolecular data - dcp protein - and on artificial data (to show the superiority of the proposed solution on better-imagined data) with positive results.

ABSTRACT. Designing efficient steel solidification methods could contribute to a sustainable future manufacturing. Current computational models, including physics-based and machine learning-based design, have not led to a robust solidification design. Predicting phase-change interface is the crucial step for steel solidification design. In the present work, we propose a simplified model for thermal radiation to be included in the phase-change equations. The proposed model forms a set of nonlinear partial differential equations and it accounts for both thermal radiation and phase change in the design. As numerical solver we implement a fully implicit time integration scheme and a Newton-type algorithm is used to deal with the nonlinear terms. Computational results are presented for two test examples of steel solidification. The findings here could be used to understand effect of thermal radiation in steel solidification. Combining the present approach with physics-based computer modeling can provide a potent tool for steel solidification design.

ABSTRACT. The boron-nitrogen (B-N) co-doped diamond with different structures have been studied by the first-principle calculations to find possible defect structures to achieve effective n-type doping. Nitrogen doped diamond itself shows the characteristics of direct bandgap, however its big gap between donor level and conduction band minimum (CBM) may contribute to its undesirable ionization energy. We found for the first time B-N co-doping as a promising method to overcome the disadvantages of N doping in diamond. B-N co-doped diamond, especially the B-N3 defect, retains the characteristics of direct band gap, and has the advantages of low ionization energy and low formation energy. The effective mass of electron/ hole of B-N co-doped diamond is less than that of pure diamond, indicating better conductivity in diamond. The N-2p states play vital role in the conduction band edge of B-N3 co-doped diamond. Hence, the B-N3 has outstanding performance and is expected to become a promising option for N-type doping in diamond.

ABSTRACT. Porting a complex scientific code, such as the E3SM land model (ELM), into a new computing architecture is challenging. The paper presents design strategies and technical approaches to develop an ELM ecosystem dynamics model with compiler directives (OpenACC) on Nvidia GPUs. Code has been refactored with advanced OpenACC features (such as deepcopy and routine directives) to reduce memory consumption and to increase the levels of parallelism through parallel loop reconstruction and new data structures. As a result, the optimized parallel implementation achieved more than a 140-time speedup (50 ms vs 7600 ms), compared to a naive implementation on a single Nvidia V100. On a fully loaded computing node with 44 CPUs and 6 GPUs, the code achieved over 3.0-time speedup, compared to the original code on CPU. Furthermore, the memory footprint of the optimized parallel implementation is 300 MB, which is around 15% of the 2.15 GB of memory consumed by a naive implementation. This study is the first effort to develop ELM components on GPUs efficiently.

ABSTRACT. QR factorization of dense matrices is a ubiquitous tool in high performance computing (HPC). From solving linear systems and least squares problems to eigenvalue problems, and singular value decompositions, the impact of a high performance QR factorization is fundamental to computer simulations and many applications. More importantly, the QR factorization on a batch of relatively small matrices has acquired a lot of attention in sparse direct solvers and low-rank approximations for Hierarchical matrices.

To address this interest and demand, we developed and present a high performance batch QR factorization for Graphics Processing Units (GPUs). We present a multi-level blocking strategy that adjusts various algorithmic designs to the size of the input matrices. We also show that following the LAPACK QR design convention, while still useful, is significantly outperformed by unconventional code structures that increase data reuse. The performance results show multi-fold speedups against the state of the art libraries on the latest GPU architectures from both NVIDIA and AMD.

ABSTRACT. Set similarity joins consists in computing all pairs of similar sets from two collections of sets. In this paper, we introduce an algorithm called MRSS-join, an extended version of our previous MRS-Join algorithm for the treatment of similarity in the trajectories. MRSS-join algorithm is based on the MapReduce computation model and a randomized redistribution approach guaranteeing perfect load balancing properties during all similarity join calculation steps while significantly reducing communication costs and the number of sets comparisons with regard to the best known algorithms based on prefix filtering. All our claims are supported by theoretical guarantees and a series of experiments that show the effectiveness of our approach in handling large datasets collections on large-scale systems.

ABSTRACT. The cross-correlation function appears in many fields with time-series data, and speeding up the computation is essential given the recent accumulation of significant amounts of data. The cross-correlationfunction can be calculated as a matrix-matrix product, and a significant speed-up can be expected utilizing Tensor Core, which is a matrix-matrix product acceleration unit of the latest NVIDIA Graphics Processing Units (GPUs). In this research, we target a new precision data type called the TensorFloat-32, which is available in the Ampere architecture. We develop a fast calculation method considering the characteristics of the cross-correlation function and TensorCore. Our method achieved a very high performance of 53.56 TFLOPS in the performance measurement assuming seismic interferometry using actual data, which is 5.97 times faster than cuBLAS, a widely used linear algebra library on NVIDIA GPUs. In addition, the accuracy of the calculation result is sufficiently high compared to the 64-bit floating-point calculation, indicating the applicability of Tensor Core operations using TensorFloat-32 for scientific calculations. Our proposed method is expected to make it possible to utilize a large amount of data more effectively in many fields.

ABSTRACT. We present an algorithm to perform trust-region-based optimization for nonlinear unconstrained problems. The method selectively uses function and gradient evaluations at different floating-point precisions to reduce the overall energy consumption, storage, and communication costs; these capabilities are increasingly important in the era of exascale computing. In particular, we are motivated by a desire to improve computational efficiency for massive climate models. We employ our method on two examples: the CUTEst test set and a large-scale data assimilation problem to recover wind fields from radar returns. Although this paper is primarily a proof of concept, we show that if implemented on appropriate hardware, the use of mixed-precision can significantly reduce the computational load compared with fixed-precision solvers.

ABSTRACT. Inverse design that directly predicts multiple structural characteristics of nanomaterials based on a set of desirable properties is essential for translating computational predictions into laboratory experiments, and eventually into products. This is challenging due to the high-dimensionality of nanomaterials data which causes an imbalance in the mapping problem, where too few properties are available to predict too many features. In this paper we use multi-target machine learning to directly map the structural features and property labels, without the need for exhaustive data sets or external optimisation, and explore the impact of more aggressive feature selection to manage the mapping function. We find that systematically reducing the dimensionality of the feature set improves the accuracy and generalisability of inverse models when interpretable importance profiles from the corresponding forward predictions are used to prioritise inclusion. This allows for a balance between accuracy and efficiency to be established on a case-by-case basis, but raises new questions about the role of domain knowledge and pragmatic preferences in feature prioritisation strategies.

ABSTRACT. We present a time-marching scheme based on the Discontinuous Petrov-Galerkin (DPG) method for linear and semi-linear transient problems. We employ an ultraweak variational formulation of the problem and we calculate the optimal test functions analytically. For semi-linear problems, we first approximate the nonlinear term by a polynomial in time employing known values of the solution from previous stages. We show that the scheme is equivalent to exponential integrators for the trace variables. Additionally, our method delivers the L^2-projection of the (linearized) solution in the element interiors and an error representation function for adaptivity in time. We prove that the error indicator is both efficient and reliable by a posteriori error estimation. We combine our time-marching scheme with Bubnov-Galerkin discretization in space. We show the performance of the scheme and adaptive strategy for 1D and 2D + time linear and semi-linear partial differential equations.

ABSTRACT. Approximating the solution of a partial differential equation (PDE) in a given quantity of interest is the main purpose of goal-oriented finite elements. The standard technique uses adaptive mesh-refinements aiming to reduce the error in such a quantity of interest.

In this work, I’m going to show a completely different approach. Namely, given a fixed (coarse) trial space, we train optimal test functions for error reduction in a fixed quantity of interest, for a whole parametric family of PDEs. The starting point will be to have a reliable training set, i.e., parameters versus quantities of interest of the solutions of the PDEs linked to those parameters. In an offline procedure, we use artificial neural networks to tune optimal test functions based on training data. The trained optimal test functions will be subsequently used in an online (Petrov-Galerkin) procedure to cheaply compute quantities of interest linked to any desired parameter, with striking accuracy.

ABSTRACT. A data-driven approach to materials design combining machine learning with molecular dynamics simulations can accelerate the discovery and development of heterogeneous catalysts such as metallic nanoparticles, using hypothetical databases of simulated nanoparticles characterised by different structural, processing, and property features. As essential catalysts, surface characteristics are particularly important for metallic nanoparticles, but they typically include a variety of atomic configurations. One of the greatest challenges in the characterisation of metallic nanoparticles is to identify surface structures relevant to catalysis that can be targeted for synthesis. Previous research has attempted to describe the surface atoms of metallic nanoparticles by their number of neighbouring atoms. While this is an intuitive approach, it fails to consider factors such as surface reconstructions and lattice variations. In this presentation, we will describe a data-driven approach to find intrinsic patterns on the surfaces of metallic nanoparticles. The method utilises iterative label spreading, an unsupervised machine learning method especially suited for clustering of small data sets with high dimensionality, which are common in materials science. Given various structural features of a nanoparticle, the model returns classes of atoms that distinguish between commonly described surface structures (e.g. certain types of surfaces, edges, corners, sub-surfaces, and sub-edges), revealing hidden patterns that could potentially relate to specific chemical reactions. We have demonstrated the application of this method on ideal and thermally relaxed palladium zonohedron nanoparticles, an important electrocatalyst with significant potential for chemical engineering.

ABSTRACT. Quite recently evolutionary algorithms (EA) turned out to be a competitor for gradient-based optimization methods in training deep neural networks. However, it is well known that simple single population evolutionary algorithms generally suffer from the problem of getting stuck in local optima. Multi-population adaptive evolutionary strategy HMS was already shown to outperform single-population EAs in multi-modal optimization and in inverse problem solving. In this paper we show a successful application of HMS in the process of deep neural network training.

ABSTRACT. Imperfect manufacturing is one of the factors affecting the performance of antenna systems. It is particularly important when design specifications are strict and leave a minimum leeway for a degradation caused by geometry or material parameter deviations from their nominal values. At the same time, conventional antenna design procedures routinely neglect to take the fabrication tolerances into account, which is mainly a result of a challenging nature of uncertainty quantification. Nevertheless, the ability to assess the effects of parameter deviations and to mitigate thereof is instrumental in achieving truly robust antenna designs. Furthermore, identifying the antenna-specific relationships between nominal requirements and tolerance immunity is essential to determine the necessary levels of fabrication accuracy, which affects both the reliability and the manufacturing costs. This paper proposes a technique for multi-criterial optimization of antenna structures oriented towards rendering a family of designs representing trade-offs between the nominal performance and the robustness. The fundamental components of our procedure are feature-based regression models constructed at the level of selected characteristic points of the antenna outputs. The trade-off designs are generated sequentially, using local search carried out for gradually relaxed nominal requirements. Numerical experiments conducted for two microstrip antennas demonstrate that the proposed algorithm is capable of yielding the performance/robustness Pareto set at the cost of only a few dozens of EM analysis of the antenna at hand per design, while ensuring reliability, as validated by means of EM-based Monte Carlo simulation.

ABSTRACT. Complex temporal epilepsy belongs to the most common type of brain disorder. Nevertheless, the wave patterns of this type of seizure, especially associated with behavioral changes, are difficult to interpret clinically. A helpful tool seems to be the statistical and time-frequency analysis of modeled epilepsy signals. The main goal of the study is the application of the Van der Pol model oscillator to study brain activity and intra-individual variability during complex temporal seizures registered in one patient. The achievement of the article is the confirmation that the statistical analysis of optimal values of three pairs of parameters of the duffing Van der Pol oscillator model enables the differentiation of the individual phases of the seizure in short-period seizure waves. In addition, the article attempts to compare the real signals recorded during the attack and modeled using frequency and time-frequency analy-sis. Similarities of power spectra and entropy samples of real and generated signals in low-frequency values are noted, and differences in higher values are explained about the clinical interpretation of the records.

ABSTRACT. Accurate information about real estate in the city, and about residential buildings in particular, is the most important resource for managing the development of the urban environment. Information about residential buildings, for example, the number of residents, is used in the inventory and digitization of the urban economy and subsequently becomes the basis of digital platforms for managing urban processes. Inventory of urban property can be carried out independently by different departments within the framework of official functions, which leads to the problem of conflicting information and missing values in urban data, in building data in particular . These problems are especially pronounced when information from different sources is combined to create centralized repositories and digital twins of the city. This leads to the need to develop approaches to filling missing values and correcting distorted information about residential buildings. As part of this work, the authors propose an approach to data imputation of residential buildings, including additional information about the environment. The analysis of the effectiveness of the approach is based on data collected for St. Petersburg (Russia)

ABSTRACT. The problem of fitting multidimensional reduced data is analyzed here. The missing interpolation knots T are substituted by Tˆ minimizing a non-linear multivariate function J_0. One of numerical schemes designed to compute such optimal knots relies on iterative scheme called Leap-Frog Algorithm. The latter is based on merging the respective generic and non-generic univariate optimizations of J^(k,i)_0. The discussion to follow analyzes the conditions enforcing unimodality of the non-generic case of J^(k,i)_0 (for special data set-up and its perturbation). This work complements already existing analysis on generic case of Leap-Frog Algorithm.

ABSTRACT. Over the recent years, utilization of numerical optimization techniques has become ubiquitous in the design of high-frequency systems, including microwave passive components. The primary reason is that the circuits become increasingly complex to meet ever-growing performance demands concerning their electrical performance, additional functionalities, as well as miniaturization. Nonetheless, as reliable evaluation of microwave device characteristics requires full-wave electromagnetic (EM) analysis, optimization procedures tend to be computationally expensive, to the extent of being prohibitive when using conventional algorithms. Accelerating EM-driven optimization is therefore a matter of practical necessity. This paper proposes a novel approach to reduced-cost gradient-based parameter tuning of passive microwave circuits with numerical derivatives. Our technique is based on restricting the finite-differentiation (FD)-based sensitivity updates to a small set of principal directions, identified as having the most significant effect on the circuit responses over the frequency bands of interest. The principal directions are found in the form of an orthonormal basis, using an auxiliary optimization process repeated before each iteration of the optimization algorithm. Extensive verification experiments conducted using a compact branch-line coupler and a dual-band power divider demonstrate up to fifty percent speedup obtained over the reference algorithm (using full FD sensitivity updates), as well as a considerable improvement over several accelerated algorithms. The computational savings are obtained with negligible degradation of the design quality.

ABSTRACT. Bi-clustering is an unsupervised data mining technique, which involves concurrent clustering of rows and columns of a two-dimensional data matrix. It has been demonstrated that bi-clustering may allow accurate and comprehensive mining of information, important for many practical applications. Numerous algorithms for data bi-clustering were proposed in the literature, based on different approaches and leading, in general, to different outputs. In this paper we propose a consensus method for combining outputs of many bi-clustering algorithms for improved quality of predictions. The proposed algorithm includes two steps. The first step, "assignment", leads to detecting groups of bi-clusters of high similarity and the second step, "trimming", results in transforming a group of similar bi-clusters into one bi-cluster of high quality. We demonstrate, on the basis of both simulated and real datasets, that using our algorithm highly improves quality of bi-clustering. We also provide an easy to use software tool, which includes implementations of several bi-clustering algorithms and our consensus method.

ABSTRACT. We consider a 2-dimensional discrete space modeled by Cellular Automata consisting of m×n cells which can be occupied by agents. There exist several types of agents which differ in their way of behavior related to their own strategy when they interact with neighbours. We assume that an interaction between agents is governed by a spatial Prisoner’s Dilemma game. Each agent participates in a number of games with his neighbors and his goal is to maximize his payoff using own strategy. Agents can change their strategies in time by replacing their own strategy by a more profitable one from its neighborhood. While agents act in such a way to maximize their incomes we study conditions of emerging collective behavior in such systems measured by the average total payoff of agents in the game or by an equivalent measure–the total number of cooperating players. These measures are the external criteria of the game, and players acting selfishly are not aware of them. We show experimentally that collective behavior in such systems can emerge if some conditions related to the game are fulfilled. We propose to introduce an income sharing mechanism to the game, giving a possibility to share incomes locally by agents. We present the results of an experimental study showing that the sharing mechanism is a distributed optimization algorithm that significantly improves the capabilities of emerging collective behavior measured by the external criterion of the game.

ABSTRACT. In this work, we propose an approach for aggregating classifiers using positional voting techniques. We propose an optimized positional voting to calculate better weights for aggregating the committee classifiers. Staring from initial weights determined by a voting algorithm the aggregating weights are optimized by a genetic algorithm. The algorithm has been evaluated on a human action dataset. We demonstrate experimentally that on SYSU 3DHOI dataset the proposed algorithm achieves superior results against recent algorithms including skeleton-based ones.

ABSTRACT. We propose and test multiple neuro-symbolic methods for sentiment analysis. They combine deep neural networks -- transformers and recurrent neural networks -- with external knowledge bases. We show that for simple models, adding information from knowledge bases significantly improves the quality of sentiment prediction in most cases. For the specific case of medium-sized sets, we obtain significant improvements over state-of-the-art transformer-based models using our proposed methods, such as Tailored KEPLER and Token Extension. We show that the cases for which the improvement occurs belong to the hard-to-learn set.

ABSTRACT. In enhanced oil recovery (EOR) processes, foam injection reduces gas mobility and increases apparent viscosity, thus increasing recovery efficiency. The quantification of uncertainty is essential in developing and evaluating mathematical models. In this work, we perform uncertainty quantification (UQ) of two-phase flow models for foam injection using the STARS model with data from a series of foam quality-scan experiments. We first performed the parameter estimation based on three datasets of foam quality-scans on Indiana limestone carbonate core samples. Then distributions of the parameters are inferred via the Markov Chain Monte Carlo method (MCMC). This approach allows propagating parametric uncertainty to the STARS apparent viscosity model. In particular, the framework for UQ allowed us to identify how the lack of experimental data affected the reliability of the calibrated models.

ABSTRACT. In this work we implement the parallelization of a method for solving fluid-structure interactions: one-field monolithic fictitious domain (MFD). In this algorithm the velocity field for solid domain is interpolated into fluid velocity field through an appropriate $L^2$ projection, then the resulting combined equations are solved simultaneously (rather than sequentially). We parallelize the finite element discretization of spatial variables for fluid governing equations and linear system solver to accelerate the computation. Our goal is to reduce the simulation time for high resolution or high dimensional fluid-structure interaction simulation, such as collision of multiple immersed solids in fluid or 3D simulations.

ABSTRACT. This work presents Direct Numerical Simulation of mass transfer in a bi-dispersed bubble swarm at high Reynolds number, by using a multiple marker level-set method. Transport equations are discretized by the finite-volume method on 3D collocated unstructured meshes. Interface capturing is performed by the unstructured conservative level-set method, whereas the multiple marker approach avoids the so-called numerical coalescence of bubbles. Pressure-velocity coupling is solved by the classical fractional-step projection method. Diffusive terms are discretized by a central difference scheme. Convective term of momentum equation, level-set equations, and mass transfer equation are discretized by unstructured flux-limiters schemes. This approach improves the numerical stability of the unstructured multiphase solver in bubbly flows with a high Reynolds number and high-density ratio. Finally, this numerical model is applied to research the effect of bubble-bubble interactions on the mass transfer in a bi-dispersed bubble swarm.

ABSTRACT. Solving general non-linear partial differential equations (PDE) precisely and efficiently has been a long-lasting challenge in the field of scientific computing. Based on the deep learning framework for solving non-linear PDEs {\itshape physics-informed neural networks} (PINN), we introduce an adaptive collocation strategy into the PINN method to improve the effectiveness and robustness of this algorithm when selecting the initial data to be trained. Instead of merely training the neural network once, multi-step discrete time models are considered when predicting the long time behaviour of solutions of the Allen-Cahn equation. Numerical results concerning solutions of the Allen-Cahn equation are presented, which demonstrate that this approach can improve the robustness of original neural networks approximation.

ABSTRACT. Fish and birds can propel themselves efficiently by acting in groups and clarifying their hydrodynamic interactions will be very useful for engineering applications. In this study, concerning the work of Becker et al., we per-formed three-dimensional unsteady simulations of an infinite array of airfoils, which is one of the models of schooling, and we clarified the structure of the flow between them. The model velocities obtained from the simulations show a good correspondence with the experimental data of Becker et al. The vortex structure created by the wing is very complicated, and it is visually clear that the vortices generated from the left and right ends of the wing also contribute to the formation of the upward and down-ward flows.

ABSTRACT. A central problem that arises in studies involving classical molecular dynamics (MD) is that they utilise protocols which do not systematically account for the chaotic nature of MD simulation. The extreme sensitivity of such simulations to their initial conditions causes the many one-off results reported to be inherently non-reproducible. Several published studies in the past few years provide evidence of extensive variability in free energies estimated using short MD simulations extending up to a few nanoseconds [1-6]. This is due to the mixing nature of the dynamics which is a necessary and sufficient condition to reach equilibrium [7-8]. Mixing implies that neighbouring trajectories, no matter how close, diverge exponentially, at a rate given by a Lyapounov exponent [7-8]. However, till date, we are not aware of any study that assessed and showed this behaviour for “long” MD simulations (of the order of a few microseconds). In this study we provide evidence for such divergence between independent simulations extending up to 10 microseconds. Our results conclusively show that MD trajectories can lead to very distinct regions of a given phase space even when they are considered “long”. Thus, results based on one-off “long” simulations are as unreliable as one-off “short” simulations and it is essential to perform ensembles in either case to quantify the uncertainty and ensure reproducibility of results.

We subject a series of five protein-ligand systems which contain important SARS-CoV-2 targets - 3CLpro, PLpro and ADRP - to long-timescale and adaptive sampling molecular dynamics simulations. By performing ensembles of ten 10-microsecond MD simulations for each system, we accurately and reproducibly determine ligand binding sites that are both crystallographically and non-crystallographically resolved, thereby discovering binding sites that can be exploited for drug discovery. We also investigate the reliability and accuracy of long-timescale trajectories. Due to the chaotic nature of molecular dynamics trajectories, individual long-timescale trajectories do not allow for accurate or reproducible elucidation of macroscopic expectation values. Upon comparing the statistical distribution of protein-ligand contact frequencies for these ten 10-microsecond trajectories, we find that over 90% of trajectories have significantly different contact frequency distributions where the associated Kolmogorov- Smirnov statistics have p values of < 0.05.

Furthermore, using a direct binding free energy calculation protocol, we determine the ligand binding free energies for each of the identified sites. We find that all long-timescale simulations lead to the computation of widely different values of binding free energy, regardless of their length. These differences range between 1.58 and 5.24 kcal/mol depending on binding site. Hence it is clear that although that is the standard way such quantities are reported, such a method cannot reliably determine the ligand binding free energy from an individual simulation to chemical accuracy (± 1 kcal/mol).

The inherent non-reproducibility of properties calculated from these individual simulations demonstrates that it is only by running ensembles of many independent trajectories that we are able to overcome the aleatoric uncertainty intrinsic to these systems in order to obtain statistically robust, reproducible and precise results.

References:

1. Sadiq, S. K.; Wright, D. W.; Kenway, O. A.; Coveney, P. V. Accurate ensemble molecular dynamics binding free energy ranking of multidrug-resistance HIV-1 protease. J. Chem. Inf. Model. 2010, 50, 890– 905, DOI: 10.1021/ci100007w 2. Genheden, S.; Ryde, U. How to obtain statistically converged MM/GBSA results. J. Comput. Chem. 2010, 31, 837– 846, DOI: 10.1002/jcc.21366 3. Wright, D. W.; Hall, B. A.; Kenway, O. A.; Jha, S.; Coveney, P. V. Computing clinically relevant binding free energies of HIV-1 protease inhibitors. J. Chem. Theory Comput. 2014, 10, 1228– 1241, DOI: 10.1021/ct4007037 4. Bhati, A. P.; Wan, S.; Wright, D. W.; Coveney, P. V. Rapid, accurate, precise, and reliable relative free energy prediction using ensemble based thermodynamic integration. J. Chem. Theory Comput. 2017, 13, 210– 222, DOI: 10.1021/acs.jctc.6b00979 5. Lawrenz, M.; Baron, R.; McCammon, J. A. Independent-trajectories thermodynamic-integration free-energy changes for biomolecular systems: Determinants of H5N1 avian influenza virus neuraminidase inhibition by peramivir. J. Chem. Theory Comput. 2009, 5, 1106– 1116, DOI: 10.1021/ct800559d 6. Bhati, A. P.; Wan, S.; Hu, Y.; Sherborne, B.; Coveney, P. V. Uncertainty quantification in alchemical free energy methods. J. Chem. Theory Comput. 2018, 14, 2867– 2880, DOI: 10.1021/acs.jctc.7b01143 7. Coveney, P. V.; Wan, S. On the calculation of equilibrium thermodynamic properties from molecular dynamics. Phys. Chem. Chem. Phys. 2016, 18, 30236-30240, DOI: 10.1039/C6CP02349E 8. Wan, S.; Sinclair, R. C.; Coveney, P. V. Uncertainty quantification in classical molecular dynamics. Phil. Trans. R. Soc. A. 2021, 379, 20200082, DOI:10.1098/rsta.2020.0082

ABSTRACT. The deep active subspace method is a neural-network based tool for the propagation of uncertainty through computational models with high-dimensional input spaces. Unlike the original active subspace method, it does not require access to the gradient of the model. It relies on an orthogonal projection matrix constructed with Gram-Schmidt orthogonalization, which is used to linearly project the (high-dimensional) input space to a low-dimensional active subspace. This matrix is incorporated into a neural network as the weight matrix of the first hidden layer, and optimized using back propagation to identify the active subspace of the input. We propose several theoretical extensions, starting with a new analytic relation for the derivatives of Gram-Schmidt vectors, which are required for back propagation. We also strengthen the connection between deep active subspaces and the original active subspace method, and study the use of vector-valued model outputs, which is difficult in the case of the original active subspace method. Additionally, we extract more traditional global sensitivity indices from the neural network to identify important inputs, and compare the resulting reduction of the input space to the dimension of the identified active subspace. Finally, we will assess the performance of the deep active subspace method on (epidemiological) problems with high dimensional input spaces, including an HIV model with 27 inputs and a COVID19 model with a 51-dimensional input space.

ABSTRACT. We present a novel MCMC algorithm, titled Multilevel Delayed Acceptance (MLDA). The algorithm is capable of sampling from the exact target distribution using a hierarchy of distributions of increasing complexity and computational cost and can be considered as an amalgam of two existing methods, namely the Delayed Acceptance (DA) MCMC of Christen & Fox (2005) and the Multilevel MCMC (MLMCMC) of Dodwell et al. (2015).

The original DA algorithm was designed to use a single coarse distribution to filter MCMC proposals before computing the fine density. We extend this approach in two ways. Vertically, by allowing any number of coarse distributions to underpin the target and horizontally, by allowing the coarse level samplers to generate extended subchains of either fixed or random lengths. The resulting algorithm is in detailed balance with the exact target distribution. We show that MLDA can be exploited for variance reduction, similarly to MLMCMC, and construct a multilevel error model that adaptively aligns the coarse distributions to the target with little additional computational cost.

The MCMC samples generated by MLDA are indeed Markov, unlike MLMCMC, where detailed balance is theoretically only ensured at infinite computational cost. However, MLDA is strictly sequential and hence suffers from resistance to parallelisation, like most MCMC algorithms. It also introduces an additional tuning parameter, namely the subsampling length for the coarse samplers. We discuss opportunities for parallelising and tuning MLDA using concepts from reinforcement learning.

ABSTRACT. Cardiovascular diseases are still responsible for many deaths worldwide, and computational models are essential tools for a better understanding of the behavior of cardiac tissue under normal and pathological situations. The microstructure of cardiac tissue is complex and formed by the preferential alignment of myocytes along their main axis with end-to-end coupling. Mathematical models of cardiac mechanics require the process of parameter estimation to produce a response consistent with experimental data and the physiological phenomenon in question. This work presents a polynomial chaos-based emulator for forward uncertainty quantification and sensitivity analysis of the Holzapfel-Ogden orthotropic constitutive model during the passive filling stage. The fiber orientation field is treated as a random field through the usage of the Karhunen-Loève (KL) expansion. The response and uncertainty of the constitutive parameters of the model considered here are also investigated. Our results show the propagated uncertainties for the end-diastolic volume and fiber strain. A global sensitivity analysis of the constitutive parameters of the complete model is also presented, evidencing the model's key parameters.

ABSTRACT. A fully automated computational tool for the study of the uncertainty in a mathematical-computational model of a heterogeneous multi-scale atomistic-continuum coupling sys-tem is implemented and tested in this project. This tool can facilitate quantitative assess-ments of the model's overall uncertainty for a given specific range of variables. The compu-tational approach here is based on the polynomial chaos expansion using projection vari-ance, a pseudo-spectral method. It also supports regression variance, a point collocation method with nested quadrature point where the random sampling method takes a diction-ary of the names of the parameters which are manually defined to vary with corresponding distributions. The tool in conjunction with an existing platform for verification, valida-tion, and uncertainty quantification offers a scientific simulation environment and data processing workflows that enables the execution of simulation and analysis tasks on a cluster or supercomputing platform with remote submission capabilities.

ABSTRACT. Solving rich vehicle routing problems is an important topic due to their numerous practical applications. Although there exist a plethora of (meta)heuristics to tackle this task, they are often heavily parameterized, and improperly tuned hyper-parameters adversely affect their performance. We exploit particle swarm optimization to select the pivotal hyper-parameters of a route minimization algorithm applied to the pickup and delivery problem with time windows. The experiments, performed on benchmark and real-life data, show that our approach automatically determines high-quality hyper-parameters of the underlying algorithm that improve its abilities and accelerate the convergence.

ABSTRACT. Nowadays, it is possible to collect precise data describing movements of public transport. Specifically, for each bus (or tram) geoposition data can be regularly collected. This includes data for all buses in Warsaw, Poland. Moreover, this data can be downloaded and analysed. In this context, one of the simplest questions is: can a model be build to represent behavior of busses, and predict their delays. This work provides initial results of our attempt to answer this question

ABSTRACT. Agent-based models (ABM) and simulations are becoming more prominent in social science. By making explicit assumptions about the behaviour of computer agents and doing simulations, mechanisms and phenomena in the society can be studied and discussed in a more formalized manner. In some settings, relevant data can be collected to verify some of the assumptions of an ABM, or the model can be verified by comparing some simulation result to data. In other settings it may not be possible to collect relevant data. Nonetheless, ABM can still be used to enrich our understanding on the research topic.For example, ABM can be constructed to study hypothetical, "what-if" scenarios in our society. While we may not be able to draw strong conclusion from these theoretical studies, they can help us reason about possible scenarios for future societies.

In particular, we are interested in modeling the far future of humanity---when human civilizations have become interstellar. We believe it is important to try to understand the possibilities of our future and be prepared for that. There has been plenty of artistic imagination on the possible future in science fiction. However, we cannot empirically study the future. Additionally, it is not clear whether current social dynamics would apply equally well to our future. As a result, we rarely see serious academic discussion on this topic. For such a hypothetical scenario, we suggest that ABM can be a viable tool to initiate meaningful discussions.

In order to build interstellar models, we only consider a scale with normal stellar objects such that we can ignore the effects of general relativity, such as universe expansion and black holes. At this scale, we still have to consider the effect of special relativity.In the context of ABM, where we the set the stage of simulation as a collection of inertial frames under the same velocity, we can simplify the theory of relativity into two core phenomena: speed of light as the upper bound of the speed of information travel, and time dilation relative to any stationary observer in the inertial frames. This also implies that we have to take care of four dimensions: three space dimension, plus one time dimensions.

Instead of building an ABM from scratch, researchers often rely on a simulation framework, such as NetLogo, to reduce the development time needed. While it is possible to build a 4D relativistic model in some existing ABM frameworks, it is not easy to enforce the relativistic effect and it can be error-prone. Therefore, we developed a simulation framework we call "Relativitization", to help social scientists to build their ABM in relativistic spacetime.

The "Relativitization" framework enforces modelers to follow a specific style of programming: define data components, determine the behavioural rules of an agent based on its environment, decide whether a mechanism is time-dilated or not, and model the active interaction between agents by command. Modelers only need to think in 3D since at every instance each agent observes only a 3D slice of spacetime, and the framework will take care of the 4D aspects. Additionally, the framework provides some utilities to help modelers implement the physics correctly. We follow standard practices in software engineering, and have designed the framework in a modular way with reusable model components, while using immutable data structure to reduce multithreading-related errors.

Our framework will lower the barrier of entry and encourage social scientists to apply their expertise to explore the interstellar future of human civilization. We hope our framework can be used to initiate meaningful and academically interesting discussions about our future.

ABSTRACT. Recently, the pseudo-Newton method was proposed to solve the problem of finding the points for which the maximal modulus of a given polynomial over the unit disk is attained. In this paper, we propose a modification of this method, which relies on the use of fractional order derivatives. The proposed modification is evaluated twofold: visually via polynomiographs coloured according to the number of iterations, and numerically by using the convergence area index, the average number of iterations and generation time of polynomiographs. The experimental results show that the fractional pseudo-Newton method for some fractional orders behaves better in comparison to the standard algorithm.

ABSTRACT. One of important areas of machine learning research is zero-shot learning. It is applied when properly labeled training data set is not available. A number of zero-shot algorithms have been proposed and experimented with. However, none of them seems to be the “overall winner”. In situations like this, it may be possible to develop a meta-classifier that would combine “best aspects” of individual classifiers and outperform all of them. In this context, the goal of this contribution is twofold. First, multiple state-of-the-art zero-shot learning methods are compared for standard benchmark datasets. Second, multiple meta-classifiers

ABSTRACT. In this paper, interval modification of degenerate parametric integral equation system (DPIES) for Laplace's equation was presented. The main purpose of such modification was to reduce the computational time necessary to solve uncertainly defined boundary problems. Such problems were defined with the uncertainly defined shape of the boundary. The shape was modeled using proposed modified directed interval arithmetic. The presented algorithm of mentioned interval modification of DPIES was implemented as a computer program. The reliability and efficiency tests were proceeded based on boundary problems modeled by Laplace’s equation. Both, obtained solutions, as well as computational time, were analyzed. As a result, the interval kernels interpolation using Lagrange polynomial caused a reduction of necessary interval arithmetic calculations, which also caused accelerations of computations.

ABSTRACT. Efficient learning of deep learning models for medical image analysis needs a large amount of labelled data. However, in many cases, we cannot build a large medical imaging dataset to train a robust model. Federated learning can help to build a shared model from multicentre data while storing the training data locally for privacy. In this paper, we propose an evaluation of deep learning-based models called CXR-FL for chest X-ray image analysis using the federated learning method. The pre-trained weights and code are publicly available (https://github.com/SanoScience/CXR-FL).

ABSTRACT. The paper presents the strategy for solving elastoplastic problems using a parametric integral equation system (PIES) and a trimming technique. It allows even complex shapes of a plastic zone to be modeled with a single surface and a set of trimming curves. New schemes for integration and approximation of solutions are developed to take into account changed requirements. However, both of them have kept their advantages. Some examples are solved, and the obtained results are compared with solutions received from other numerical methods.

ABSTRACT. We employ a variational splitting for Crank-Nicolson method and parabolic partial differential equations. We exploit the tensor-product structure of the discretization in space (mesh and basis functions) to derive an update strategy that has a linear cost with respect to the total number of degrees of freedom for three-dimensional parabolic problems. We present a method to extend our linear computational cost variational splitting implicit solver for the case with variable material data parameters. In this case, the material data can vary locally with different quadrature point. We illustrate this method with numerical results of Penner bioheat equation modeling the heating of the human head as a result of the cellphone antenna radiation.

ABSTRACT. Nowadays, a wide range of critical services relies on Internet of Things (IoT) devices. Nevertheless, they often lack proper security, becoming the gateway to attack the whole system. IoT security protocols are often based on stream ciphers, where pseudo-random number generators (PRNGs) are an essential part of them. In this work, a family of strong ciphers difficult to be analyzed by standard methods is described. In addition, we will discuss an innovative technique of sequence decomposition (binomial representation) and present a novel algorithm based on Hadamard matrices to evaluate the strength (unpredictability) of binary sequences, a key part of the IoT security stack.

ABSTRACT. Error correction is wide and well elaborated area of quan-tum information theory. There are many works and methods of thisarea. Those methods, however, demand additional resources, like quan-tum gates, qubits or time. We have observed, in statistical sense, that thequbit’s error in real quantum computers, once calibrated doesn’t changemuch until next one. Then being so, for quantum sampling based com-putations, one can determine the correction experimentally and use ituntil the next calibration, without a need of utilize additional resources.In this work we present the method of determining such a correction andapplying it to practical quantum-sampling algorithms.Quantum sampling is the method, which we deliberately decline to ob-tain one deterministic result of one-shot computation in. Instead of that,we provide a number of same experiments, with the same initializationstate, the same evolution operator and the same measurement basis.Then we observe the probability distribution function (PDF) thus ob-tained, which is considered as the final result of computation. We haveobserved and experimentally proved in this work, that error of this prob-ability distribution is correlated with the local quantum phase of qubitsinvolved in computations. Hence we are able to create aPhase DistortionUnraveling(PDU) function for each qubit and for whole system as well,that depends on this phase. Briefly, the final result after correction is thesum of PDF and PDU.

ABSTRACT. The classification of music genres is essential due to millions of songs in online databases. It would be nearly impossible or very costly to do this job manually. That is why there is a need to create robust and efficient methods that automatically help to do this task. In this paper, music genre recognition is implemented by exploiting the potential of wide ensembles of neural network classifiers. Creating infrequently used types of ensembles is a main contribution of authors in the development of automatic recognition of the musical genre. The paper shows how it can be done in a relatively quick and straightforward manner. The presented method can be implemented in many other use cases.

ABSTRACT. We developed and validated a language-agnostic method for sentiment analysis. Cross-language experiments carried out on the new MultiEmo dataset with texts in 11 languages proved that LaBSE embeddings with an additional attention layer implemented in the biLSTM architecture outperformed other methods in most cases. Therefore, it has been implemented in our online web service.

ABSTRACT. This paper presents an algorithm for fitting the boundary geometry with Bé-zier curves in the parametric integral equation system (PIES). The algorithm determines the coordinates of control points by minimizing the distance be-tween the constructed curves and contour points on the boundary. The min-imization is done with the Adam optimizer that uses the gradient of the ob-jective function calculated by automatic differentiation (AD). Automatic dif-ferentiation eliminates error-prone manual routines to evaluate symbolic de-rivatives. The algorithm automatically adjusts to the actual number of curves and their degree. The results of the presented tests show the high accuracy and scalability of the proposed approach. Finally, we demonstrate that the resulting boundary may be directly used by the parametric integral equation system (PIES) to solve the boundary value problem in 2D governed by the Laplace equation.

ABSTRACT. Kubernetes has gained extreme popularity as a cloud-native platform for distributed applications. However, scientific computations which typically consist of a large number of jobs -- such as scientific workflows -- are not typical workloads for which Kubernetes was designed. In this paper, we investigate the problem of execution of scientific workflows on a Kubernetes cluster with particular focus on autoscaling, i.e. adjusting the computing infrastructure to the current resource demands. We discuss alternative models for execution of scientific workflows in Kubernetes and propose a solution for auto-scaling that takes advantage of the known workflow structure to improve scaling decisions by predicting resource demands for the near future. Such a predictive autoscaling policy is experimentally evaluated and compared to a regular reactive policy where only the current demand is taken into account. The experimental evaluation is done using the HyperFlow workflow management systems running five simultaneous instances of the Montage workflow on a Kubernetes cluster deployed in the Google Cloud Platform. The results indicate that the predictive policy allows achieving better elasticity and execution time, while reducing monetary cost.

ABSTRACT. Our work is positioned at the intersection of game data science and affective gaming. We propose an original method to evaluate the quality of a selected class of video games based of emotional reactions of players. The developers of free to play mobile games ask themselves why some games are more profitable (MP games) and other less (LP games). Since the success of these games depends on the player's attachment to the game, a naive and general, but intuitively convincing hypothesis is often put forward -- MP games evoke more positive emotions, and hence crucially, are sustainably engaging. Our main hypothesis is: test players who can clearly distinguish between MP game and LP game in relatively short test sessions are more reliable and interesting to keep track of their emotions. From a random group of test players, we selected players who had such abilities. We analyzed their affective spectra and got a fairly clear confirmation that the selected players showed more positive and less negative emotions in MP games than in LP ones during the test sessions. We can reasonably expect these players to be focused on playing in the test session, and their emotions may really indicate the strengths of MP games or weaknesses of LP games. We present the results of the experimental evaluation of our method which was conducted with a close cooperation with one of the leading game companies in Poland.

ABSTRACT. In the context of humanitarian support for forcibly displaced persons, camps play an important role in protecting people and ensuring their survival and health. A challenge in this regard is to find optimal locations for establishing a new asylum-seeker/unrecognized refugee or IDPs (internally displaced persons) camp. In this paper we formulate this problem as an instantiation of the well-known facility location problem (FLP) with three objectives to be optimized. In particular, we show that AI techniques and migration simulations can be used to provide decision support on camp placement. 

ABSTRACT. Numerous words exhibit varying meanings in different contexts of use. Despite ongoing development of neural methods in distributional semantics e.g. contextual embeddings, the automated mapping of word occurrences to the respective senses in a semantic lexicon such as lexical semantic network may facilitate many applications of Natural Language Processing. On the other hand, neural language models and multilingual transfer learning techniques have been proved to be very effective when applied to downstream tasks in low-resource languages. In this paper we introduce a comprehensive evaluation benchmark for Polish Word Sense Disambiguation task. As far as we know, our work is a first attempt to standardise existing sense annotated data for WSD in Polish. We also follow the recent trends of neural WSD and we zero-shot transfer learning models as well as hybrid WSD architectures combining lexico-semantic networks with neural context encoders. Finally, we investigate the impact of bilingual training on WSD performance. The bilingual model obtains new state of the art performance in Polish WSD task.

ABSTRACT. Finding the~ground state of the~Ising spin-glass is an important and challenging problem (NP-hard, in fact) in condensed matter physics. However, its applications spread far beyond physic due to its deep relation to various combinatorial optimization problems, such as travelling salesman or protein folding. Sophisticated and promising new methods for solving Ising instances rely on quantum resources. In particular, quantum annealing is a quantum computation paradigm, that is especially well suited for Quadratic Unconstrained Binary Optimization (QUBO). Nevertheless, commercially available quantum annealers (i.e., D-Wave) are prone to various errors, and their ability to find low energetic states (corresponding to solutions of superior quality) is limited. This naturally calls for a post-processing procedure to correct errors (capable of lowering the~energy found by the~annealer). As a proof-of-concept, this work combines the~recent ideas revolving around the~DIRAC architecture with the~Chimera topology and applies them in a real-world setting as an error-correcting scheme for quantum annealers. Our preliminary results show how to correct states output by quantum annealers using reinforcement learning. Such an approach exhibits excellent scalability, as it can be trained on small instances and deployed for large ones. However, its performance on the chimera graph is still inferior to a typical algorithm one could incorporate in this context, e.g., simulated annealing.

ABSTRACT. With the advance of experimental techniques insight into cognitive processes of memorising and retrieving information, and memory distortions has become pos- sible. One such technique being functional magnetic resonance imaging (fMRI). In recent years, the functional activations have been intensely analysed by a range of machine learning [1] and deep learning [4] methods to study brain disor- ders. However, to study cognitive processes one must use repeatable tasks which make fMRI data notoriously difficult to analyse due to its very low temporal resolution. In this work, we apply several linear and non-linear classification methods to fMRI signals from the short-term memory experiment [3]. The experiment con- sisted of two visual verbal tasks (based on lists of semantically or phonetically associated words), two non-verbal tasks (pictures of similar objects), and spon- taneous brain activity (resting state). The methods included among several oth- ers: Quadratic Discriminant Analysis (QDA), Random Forests, hyperparameter tuned Light Gradient Boosting [2] (LGM), and ResNets he2016deep of several depths. With these methods we automatically classify very short segments of brain activity (1-6 samples, corresponding to 2-11 seconds) during information retrieval/recognition into stimuli types (2, 3, 4, or 5 classes). We show that the best classifiers reach F1-scores up to .834 for 2-class and up to .603 for 5-class problem. The nonlinear classifiers (such as QDA, LGBM or ResNets) clearly beat linear ones, but none of them is universally best in our experiment. Most interestingly, our tests showed that information crucial for producing good classification results is localised in a small number of ROIs. Presented findings increase the efficiency of fMRI signal classification in cognitive experiments and give rise to understanding of the cognition process related to short-term memory performance and distortions.

ABSTRACT. Agent-based models frequently make use of scaling techniques to render the simulated samples of population more tractable. This work presents a simulation of the spread of disease and assesses the extent to which simulations vary relative to the samples upon which they are based. It is determined that in particular, geographical aspects of the spread of disease are best capture by larger, higher fidelity models of a population.

ABSTRACT. The processes of neurodegeneration related to Alzheimer’s disease (AD) begin several decades before the first symptoms. We have used granular computing rules (rough set theory) to classify cognitive data from BIOCARD study that have been started over 20 years ago with 354 normal subjects. Patients were evaluated every year by team of neuropsychologists and neurologists and classified as normal, with MCI (mild cognitive im-pairments), or with dementia. As the decision attribute we have used CDRSUM (Clinical Dementia Rating Sum of Boxes) as more quantitative measure than above classification. Based on 150 stable subjects with different stages of AD we have found rules (granules) that classify cognitive attributes with disease stages (CDRSUM). By applying these rules to normal (CDRSUM=0) 21 subjects we have predicted that one subject might get mild dementia (CDRSUM > 4.5), and 10 other might get questionable cognitive impairment (CDRSUM>0.75). AI methods can find, invisible for neuropsychologists, patterns in cognitive attributes of normal subjects that might indicate their pre-dementia stage.

ABSTRACT. We focus on Isogeometric Analysis (IGA) approximations of Partial Differential Equations (PDEs) solutions. We consider linear combinations of high-order and continuity base functions utilized by IGA. Instead of using the Deep Neural Network (DNN), which is the concatenation of linear operators and activation functions, to approximate the solutions of PDEs, we employ the linear combination of higher-order and continuity base functions, as employed by IGA. In this paper, we compare two methods. The first method trains different DNN for each coefficient of the linear computations. The second method trains one DNN for all coefficients of the linear combination. We illustrate our consideration using model elliptic problems.

ABSTRACT. Fetal arrhythmia is an abnormal heart rhythm caused by a problem in the fetus's heart's electrical system. Monitoring fetal ECG is vital to delivering useful information regarding the fetus's condition. Acute fetal arrhythmia may result in cardiac failure or death. Thus, the early detection of fetal arrhythmia is important. Current approaches use several electrodes to acquire abdomen ECG from the mother which causes discomfort. Moreover, ECG signals acquired are extremely noisy and have artifacts that occur from breathing, muscle contraction which harden ECG extraction. In this study, a machine learning framework for fetal arrhythmia detection is proposed. The proposed framework uses only a single abdomen ECG. It employs multiple filtering techniques to remove noise and artifacts. It also extracts several significant features from multiple domains including (time and time-frequency features). Finally, it utilizes four machine learning classifiers to detect arrhythmia. The highest accuracy of 93.12% is achieved using boosted decision tree classifier. The performance of the proposed method shows its competing ability compared to other methods.

ABSTRACT. When working with a mix of qualitative and quantitative data for constructing ABMs, a structured framework ensures proper information editing and model validation. For this, we propose FREIDA (Framework for expert-informed data-driven agent-based-models).

With this proposed framework, an ABM can be constructed in a reproducible manner and taking into account both qualitative (e.g., case files, interviews, ODD+D) and quantitative data (e.g., databases, experimental outcomes, and literature), and providing a guide to cross-referencing, editing, and verifying the data and resulting simulation. FREIDA is not only a continuation of the existing ODD+D framework, but also extends the modelling process to the verification stage with verification statements, sensitivity analysis, and validation (e.g., using a new set of case files). This ensures model validity in high-stakes modelling cases despite scarce availability of data, such as when modelling the replacement behaviour within criminal networks.

The framework improves on previously existing methodologies for modelling and documentation. We expect to improve the framework further by testing it in more cases in cooperation with the Dutch Police, as well as making FREIDA truly iterative and thus completing the empirical cycle.

ABSTRACT. This talk will introduce a productive blend of AI and physics-based (PB) approaches which can be used to create digital twins of the human body to deliver a new generation of truly personalised and predictive medicine. An iterative cycle is used in Big AI – the synergy of ML and PB, where AI hypotheses are tested in physics-based simulations, and the results of PB modelling are used to train AI. A workflow has been constructed for Big AI with ensemble approaches, which overcomes the issue of variability in predictions and leads to true, and commonly non-Gaussian, statistics. Big AI could be used in many ways: some examples including tracking the spread of the virus in pandemics, predicting accurate drug binding affinities to their target proteins, accelerating drug discovery and vaccine design, and proposing drugs or drug combinations for the treatment of diseases.

ABSTRACT. Parallel three-dimensional (3D) cellular automaton models of tumor growth can efficiently model tumor morphology over many length and time-scales. Here, we extended an existing two-dimensional (2D) model of tumor growth to model how tumor morphology could change over time and verified the 3D model with the initial 2D model on a per-slice level. However, increasing the dimensionality of the model imposes constraints on memory and time-to-solution that could quickly become intractable when simulating long temporal durations. Parallelizing such models would not only enable larger tumors to be investigated but also pave way for coupling with treatment models. We parallelized the 3D growth model using N-body and lattice halo exchange schemes, and further optimized the implementation to adaptively exchange information based on the state of cell expansion. We demonstrate a factor of 20x speedup compared to the serial model when running on 340 cores of Stampede2’s Knight’s Landing compute nodes. This proof-of-concept study demonstrated that parallel 3D models could enable exploration of large problem and parameter spaces at tractable run times.

ABSTRACT. Simulations of the microvasculature can elucidate the effects of various blood flow parameters on micro-scale cellular and fluid phenomena. At this scale, the non-Newtonian behavior of blood requires the use of explicit cell models, which are necessary for capturing the full dynamics of cell motion and interactions. Over the last few decades, fluid-structure interaction models have emerged as a method to accurately capture the behavior of deformable cells in the blood. However, as computational power increases and systems with millions of red blood cells can be simulated, it is important to note that varying spatial distributions of cells may affect simulation outcomes. Since a single simulation may not represent the ensemble behavior, many different configurations may need to be sampled to adequately assess the entire collection of potential cell arrangements. In order to determine both the number of distributions needed and which ones to run, we must first establish methods to identify well-generated, randomly-placed cell distributions and to quantify distinct cell configurations. In this work, we utilize metrics to assess 1) the presence of any underlying structure to the initial cell distribution and 2) similarity between cell configurations. We propose the use of the radial distribution function to identify long-range structure in a cell configuration and apply it to a randomly-distributed and structured set of red blood cells. To quantify spatial similarity between two configurations, we make use of the Jaccard index, and characterize sets of red blood cell and sphere initializations.

ABSTRACT. Ductal carcinoma in-situ (DCIS) presents a risk of transformation to malignant intraductal carcinoma (IDC) of the breast. Three tumor suppressor genes RB, BRCA1 and TP53 are critical in curtailing the progress of DCIS to IDC. The complex transition process from DCIS to IDC involves acquisition of intracellular genomic aberrations and consequent changes in phenotypic characteristics and protein expression level of the cells. The spatiotemporal dynamics associated with breech of epithelial basement membrane and subsequent invasion of stromal tissues during the transition is less understood. We explore the emergence of invasive behavior in benign tumors, emanating from altered expression levels of the three critical genes. A multiscale mechanistic model based on Glazier-Graner-Hogeweg method-based modelling (GGH) is used to unravel the phenotypical and biophysical dynamics promoting the invasive nature of DCIS. Ductal morphologies including comedo, hyperplasia and DCIS, evolve spontaneously from the interplay between the gene activity parameters in the simulations. The spatiotemporal model elucidates the cause and effect relationship between cell-level biological signaling and tissue-level biophysical response in the ductal microenvironment. The model predicts that BRCA1 mutations will act as a facilitator for DCIS to IDC transitions while mutations in RB act as initiator of such transitions.

ABSTRACT. High Content Screening allows for a complex cell analysis by combining fluorescent microscopy with the capability to automatically create a large number of images. Example of such cell analysis is examination of cell morphology under influence of a compound. Nevertheless, classical approaches bring the need for manual labeling of cell examples in order to train a machine learning model. Such methods are time- and resource-consuming. To accelerate the analysis of HCS data, we propose a new self-supervised model for cell classification: Self-Supervised Multiple Instance Learning (SSMIL). Our model merges Contrastive Learning with Multiple Instance Learning to analyze images with weak labels. We test SSMIL using our own dataset of microglia cells that present different morphology due to compound-induced inflammation. Representation provided by our model obtains results comparable to supervised methods proving feasibility of the method and opening the path for future experiments using both HCS and other types of medical images.

ABSTRACT. In recent years, 5G network systems start to offer communication infrastructure for Internet of Things (IoT) applications, especially for health care service providers. In smart health care systems, edge computing enabled Internet of Medical Things (IoMT) is an innovative technology to provide online health care monitoring facility to patients. Here, energy consumption, along with ex-tending the lifespan of biosensor network, is a key concern. In this contribution, a Chicken Swarm Optimization algorithm, based on Energy Efficient Multi-ob-jective clustering scheme is applied in the context of IoMT system. An effective fitness function is designed for cluster head selection, using multiple objectives, such as residual energy, queuing delay, communication cost, link quality and node centrality. Simulated outcomes of the proposed scheme are compared with the existing schemes in terms of parameters such as cluster formation time, energy consumption, network lifetime, throughput and propagation delay.

ABSTRACT. The Partitioned Global Address Space (PGAS) model is well suited for executing irregular applications on cluster-based systems, due to its efficient support for short, one-sided messages. Separately, the actor model has been gaining popularity as a productive asynchronous message-passing approach for distributed objects in enterprise and cloud computing platforms, typically implemented in languages such as Erlang, Scala or Rust. To the best of our knowledge, there has been no past work on using the actor model to deliver both productivity and scalability to PGAS applications on clusters.

In this paper, we introduce a new programming system for PGAS applications, in which point-to-point remote operations can be expressed as fine-grained asynchronous actor messages. In this approach, the programmer does not need to worry about programming complexities related to message aggregation and termination detection. Our approach can also be viewed as extending the classical Bulk Synchronous Parallelism model with fine-grained asynchronous communications within a phase or superstep. We believe that our approach offers a desirable point in the productivity-performance space for PGAS applications, with more scalable performance and higher productivity relative to past approaches. Specifically, for seven irregular mini-applications from the Bale benchmark suite executed using 2048 cores in the NERSC Cori system, our approach shows geometric mean performance improvements of ≥ 20 times relative to standard PGAS versions (UPC and OpenSHMEM) while maintaining comparable productivity to those versions.

ABSTRACT. For appropriate workloads, partially ordered message delivery can greatly reduce message latency. For example, updates to screens (e.g., remote desktops, VNC) may not have to be totally ordered with respect to different regions of the screen, but ordered with respect to updates to the same region. Similarly, updates to disjoint regions of a file can be applied in any order, as long as updates (or reads) to the same region of the file are ordered in a consistent way, per consistency models. Therefore, we introduce the concept of consistency fences (CFs), inspired by memory fences from data consistency models, as a mechanism to control, specify, and reason about partial orders. If messages are lost on a network, partial ordering via CFs provides a framework to tolerate the latency associated with retransmission, for key workloads. In a set of simple experiments, based on screen update workloads, we show the latency benefits of partial ordering with CFs. We also show how forward error-correction can be combined with CFs and partial ordering to reduce cumulative latency (represented as a cumulative distribution function), as compared to total ordering of messages.

ABSTRACT. In the paper we present performance-energy trade-off investigation of training Deep Convolutional Neural Networks for image recognition. Several representative and widely adopted network models, such as Alexnet, VGG-19, Inception V3, Inception V4, Resnet50 and Resnet152 were tested using systems with Nvidia Quadro RTX 6000 as well as Nvidia V100 GPUs. Using GPU power capping we found other than default configurations minimizing three various metrics: energy (E), energy-delay product (EDP) as well as energy-delay sum (EDS) which resulted in considerable energy savings, with a low to medium performance loss for EDP and EDS. Specifically, for Quadro 6000 and minimization of E we obtained energy savings of 28.5%-32.5%, for EDP 25%-28% of energy was saved with average 4.5%-15.4% performance loss, for EDS (k=2) 22%-27% of energy was saved with 4.5%-13.8% performance loss. For V100 we found average energy savings of 24%-33%, for EDP energy savings of 23%-27% with corresponding performance loss of 13%-21% and for EDS (k=2) 23.5%-27.3% of energy was saved with performance loss of 4.5%-13.8%.

ABSTRACT. With the development of distributed stream processing systems, elastic resource allocation has become a powerful means to deal with the fluctuating data stream. The existing methods either focus on a single operator or only consider the static correlation between operators to perform elastic scaling. However, they ignore the dynamic correlation between operators in data stream processing applications, which leads to lagging and inaccuracy resource allocation, increasing processing latency. To address these issues, we propose an elastic resource allocation method, which is based on the dynamic perception of operator influence domain, to perform resource allocation dynamically and in advance. The experimental results show that compared with the existing methods, our method not only guarantees that the end-to-end latency meets QoS requirements but also reduces resource utilization.

ABSTRACT. Data-flow computing is a natural and convenient paradigm for expressing parallelism. This is particularly true for tools that automatically extract the data dependencies among the tasks while allowing to exploit both distributed and shared memory parallelism. This is the case of UPC++ DepSpawn, a new task-based library developed on UPC++ (Unified Parallel C++), a library for parallel computing on a Partitioned Global Address Space (PGAS) environment, and the well-known Intel TBB (Threading Building Blocks) library for multithreading. In this paper we present and evaluate the evolution of this library after changing its engine for shared memory parallelism and adapting it to the newest version of UPC++, which differs very strongly from the original version on which UPC++ DepSpawn was developed. As we will see, while keeping the same high level of programmability, the new version is on average 9.3% faster than the old one, the maximum speedup being 66.3%.

ABSTRACT. Neurons exhibit striking complexity and diversity in their geometry, which is essential for neuronal functions and biochemical signal transmission. However, it also brings challenges to mediate intracellular material transport since most essential materials for neurons have to experience long-distance transport along axons and dendrites after synthesis in the cell body. In particular, the neuron relies heavily on molecular motors for the fast transport of various materials along the cytoskeletal structure like microtubules (MTs). The disruption of this long-distance transport can induce neurological and neurodegenerative diseases like Huntington’s, Parkinson’s, and Alzheimer’s disease. Therefore, it is essential to study the intracellular transport process in neurons. Here, we present to simulate the intracellular material transport within complex neuron geometries using several methods including isogeometric analysis (IGA), deep learning (DL), and PDE-constrained optimization (PDE-CO).We obtain the velocity and concentration results under normal and traffic jam transport conditions within 3D neuron geometries. Moreover, we develop a DL-based surrogate model to improve the computational efficiency of the IGA simulation. Our results provide key insights into how material transport in neurons is mediated by their complex geometry and MT distribution.

ABSTRACT. The rate of penetration (ROP) is a key performance indicator in the oil and gas drilling industry as it directly translates to cost savings and emission reductions. A prerequisite for a drilling optimization algorithm is a predictive model that provides expected ROP values in response to surface drilling parameters and formation properties. The high predictive capability of current machine-learning models comes at the cost of excessive data requirements, poor generalization, and extensive computation requirements. These practical issues hinder ROP models for field deployment. Here we address these issues through transfer learning. Simulated and real data from the Volve field were used to pre-train models. Subsequently, these models were fine-tuned with varying retraining data percentages from other Volve wells and Marcellus Shale wells.

Four out of the five test cases indicate that retraining the base model would always produce a model with lower mean absolute error than training an entirely new model or using the base model without retraining. One was on par with the traditional approach. Transfer learning allowed to reduce the training data requirement from a typical 70 percent down to just 10 percent. In addition, transfer learning reduced computational costs and training time. Finally, results showed that simulated data could be used in the absence of real data or in combination with real data to train a model without trading off model’s predictive capability.

ABSTRACT. The main idea of our Goal-Oriented Adaptive strategy is based on performing global and uniform h- or p-refinement (for h- and p-adaptivity, respectively) followed by a coarsening step, where some basis functions are removed according to their estimated importance. Many Goal-Oriented Adaptive (GOA) strategies represent the error in a Quantity of Interest (QoI) in terms of the bilinear form and the solution of the direct and adjoint problems. However, this is unfeasible when solving indefinite or non-symmetric problems since symmetric and positive definite forms are needed to define the inner product that guides the refinements. In this work, we provide a Goal-Oriented Adaptive (h- or p-) strategy whose error in the QoI is represented in another bilinear symmetric positive definite form than the one given by the adjoint problem. For that purpose, our Finite Element implementation employs a multi-level hierarchical data structure that imposes Dirichlet homogeneous nodes to avoid the so-called hanging nodes. We illustrate the convergence of the proposed approach for 1D Helmholtz and convection-dominated problems.

ABSTRACT. Solving Partial Differential Equations (PDEs) has been a long-standing challenge in numerical analysis, engineering, and computation. Traditional numerical methods have been widely developed and studied in the last decades, and strong contributions have been found about their convergence. For example, in a weak formulation setting, \begin{equation}\label{PG} \displaystyle{ \left\{ \begin{aligned} \mbox{Find}\;u^*&\in U\;\mbox{such that}\\ b(u^*,v)&=l(v),\;\forall v\in V, \end{aligned} \right.} \end{equation} the following optimization process solves the problem: \begin{equation}\label{minsup_residual} u^* = \arg\min_{u\in U} \sup_{0\neq v\in V} \frac{|b(u, v)-l(v)|}{||v||_V}, \end{equation} where $U$ and $V$ are the trial and test spaces, respectively, $b$ is a bilinear form, $l$ is a linear functional, and $||\cdot||_{V}$ is a norm induced in $V$. Note that \eqref{minsup_residual} is an optimization problem where the residual of \eqref{PG} is minimized in the dual norm of the test space $V$ \cite{demkowicz2014overview}.

Neural Networks have demonstrated their great power in recent years on solving PDEs. In particular, \cite{Bao_2020} and \cite{ZANG} propose to solve \eqref{minsup_residual} combining two networks adversarially: one to approximate the trial solution, and another one for approximating the test maximizer. However, they select suboptimal norms for the optimization.

In our work, we revisit the Weak Adversarial Networks, we make an extensive review of the related theory to improve their problem setting, and we propose enhanced optimization strategies. Moreover, we extend their proposal to a wider class of problems and formulations, including strong and ultra-weak formulations.

ABSTRACT. Physics-informed neural networks allow models to be trained by physical laws described by general nonlinear partial differential equations. However, traditional architectures struggle to solve more challenging time-dependent problems due to their architectural nature. In this work, we present a novel physics-informed framework for solving time-dependent partial differential equations. Using only the governing differential equations and problem initial and boundary conditions, we generate a latent representation of the problem’s spatio-temporal dynamics. Our model utilizes discrete cosine transforms to encode spatial frequencies and re-current neural networks to process the time evolution. This efficiently and flexibly produces a compressed representation which is used for additional conditioning of physics-informed models. We show experimental results on the Taylor-Green vortex solution to the Navier-Stokes equations. Our proposed model achieves state-of-the-art performance on the Taylor-Green vortex relative to other physics-informed baseline models.

ABSTRACT. Modern microwave design has become heavily reliant on full-wave electromagnetic (EM) simulation tools, which are necessary for accurate evaluation of microwave components. Consequently, it is also indispensable for their development, especially the adjustment of geometry parameters, oriented towards performance improvement. However, EM-driven optimization procedures incur considerable computational expenses, which may become impractical even in the case of local tuning, and prohibitive whenever global search is vital (e.g., multi-model tasks, simulation-based miniaturization, circuit re-design within extended ranges of operating frequencies). This work presents a novel approach to a computation-ally-efficient globalized parameter tuning of microwave components. Our framework employs the response feature technology, along with the inverse surrogate models. The latter permit low-cost exploration of the parameter space, and identification of the most advantageous regions that contain designs featuring performance parameters sufficiently close to the assumed target. The initial parameter vectors rendered in such a way undergo then local, gradient-based tuning. The in-corporation of response features allows for constructing the inverse model using small training data sets due to simple (weakly-nonlinear) relationships between the operating parameters and dimensions of the circuit under design. Global optimization of the two microstrip components (a coupler and a power divider) is carried out for the sake of verification. The results demonstrate global search capability, excellent success rate, as well as remarkable efficiency with the average optimization cost of about a hundred of EM simulations of the circuit necessary to conclude the search process.

ABSTRACT. Soil bacteria play a fundamental role in plant growth. This paper focuses on developing and testing some techniques designed to identify automatically such microorganisms. More specifically, the recognition performed here deals with the specific five genera of soil bacteria. Their microscopic images are classified with machine learning methods using shape and image texture descriptors. Feature determination based on shape relies on interpolation and curvature estimation whereas feature recognition based on image texture resorts to the spatial relationships between chrominance and luminance of pixels using co-occurrence matrices. From the variety of applied modelling methods the best reported result amounts to 97% of accuracy. This outcome is obtained upon incorporating the set of features from both groups and subsequently merging classification and feature selection methods: Extreme Learning Machine - Radial Basis Function with Sparse Multinomial Logistic Regression with Bayesian Regularization and k-Nearest Neighbors classifier with Fast Correlation Based Filter. The optimal parameters involved in merged classifiers are obtained upon computational testing and simulation.

ABSTRACT. The study introduces a System Dynamics Modelling (SDM) approach with a remote Group Model Building (GMB) component used as an answer to the climate problems of cities related to severe air pollution. The main objective of our research is twofold: to identify the factors and mechanisms that are key for the elimination of fossil-fuel boilers (FFBs) in Poland and to provide a system understanding of the underlying causal relationships and their implications to facilitate eradication of FFBs. The first phase of modelling process is presented: action research represented by remote GMB workshops, attended by 14 participants from Poland and Norway. The workshops’ results help to identify the key variables, capture the main causal relationships between them, helping to develop the first outline of the causal structure to underlie the simulation model. Despite holding the workshops during social isolation, they proved to be good educational facilitation tool for a key element of the modelling process in the SD methodology - the collection of input data for model building, as well as raising participants' awareness of climate change issues. The novelty of the study lies in the application of a proven tool such as dynamic systems modelling to applied research for sustainable development. In particular, using modelling and simulation approaches to support development and practical implementation of the energy transition policies involving the elimination of FFBs and replacement with renewable energy sources. To the best of our knowledge, only a few such studies are being conducted.

ABSTRACT. Large concentrations of particulate matter in residential areas are related to the lack of vertical movements of air masses. Their disappearance is associated with the occurrence of the most common ground temperature inversion, which inhibits the natural air convection. As a result, air layers separated by a temperature inversion layer are formed, which practically do not interact with each other. Therefore, to reduce the concentration of particulate matter, mixing of air layers should be forced, or natural processes should be restored. For this purpose, it was proposed to generate shock waves of high pressure in the vertical direction to mix the polluted air and break the inversion layer locally. This paper performs fast isogeometric analysis simulations of the thermal inversion and pollution removal process. We employ a linear computational cost solver using Kronecker product-based factorization. We compare our numerical simulations to an experiment performed with an anti-hail cannon in a highly polluted city of Krak\'ow, Poland.

ABSTRACT. Modelling of large scale data series is of significant importance in fields such as astrophysics and finance. The continuous increase in available data requires new computational approaches such as the use of multicore processors and accelerators. Recently, a novel time series modelling and forecasting method was proposed, based on a recursively updated pseudoinverse matrix which enhances parsimony by enabling assessment of basis functions, before inclusion into the final model. Herewith, a novel GPU (Graphics Processing Unit) accelerated matrix based auto-regressive variant is presented, which utilizes lagged versions of a time series and interactions between them to form a model. The original approach is reviewed and a matrix multiplication based variant is proposed. The GPU accelerated and hybrid parallel versions are introduced, utilizing single and mixed precision arithmetic to increase GPU performance. Discussions around performance improvement and high order interactions are given. A block processing approach is also introduced to reduce memory requirements for the accelerator. Furthermore, the inclusion of constraints in the computation of weights, corresponding to the basis functions, with respect to the parallel implementation are discussed. The approach is assessed in a series of model problems and discussions are provided.

ABSTRACT. The widely addressed topic of ontology alignment to this day contains several open research questions that remain either unanswered or only vaguely tackled. One of them is designating alignments of concept instances, which according to the literature are addressed in a handful of publications. Therefore, in this paper we propose a formal framework based on fuzzy logic that can be used to determine such mappings. We provide several similarity functions and a set of inference rules for combining them. The approach has been experimentally verified using widely accepted datasets provided by the Ontology Alignment Evaluation Initiative, yielding promising results.

ABSTRACT. The goal of the paper is to propose a new version of the Weighted Ensemble with one-class Classification and Over-sampling and Instance selection (WECOI) algorithm. This paper describes WECOI and presents the alternative approach for over-sampling, which is based on a selection of reference instances from produced clusters. This approach is flexible on applied clustering methods; however, the similarity-based clustering algorithm has been proposed as a core. For clustering, different methods may also be applied. The proposed approach has been validated experimentally using different clustering methods and shows how the clustering technique may influence synthetic instance generation and the performance of WECOI. The WECOI approach has also been compared with other algorithms dedicated to learning from imbalanced data streams. The computational experiment was carried out using several selected benchmark datasets. The computational experiment results are presented and discussed.

ABSTRACT. As social networks continue to grow in popularity, it is essential to understand what can be learned about private attributes of social-network users by mining social-network data. Previous work focused on the inference of time-invariant attributes such as personality traits. By contrast, we ask to what extent Dynamic Bayesian Networks can infer \textit{time-varying behavioral intentions} such as vaccination intentions or job-searching intentions. Knowledge of such intentions has great potential to improve the design of recommendation systems, ad-targeting mechanisms, and other social and commercial endeavors.

The contribution of this paper is twofold. First, we propose a novel methodology for intention inference. We design modular bayesian-network models that are able to capture the evolving nature of the human decision-making process by combining data and priors from multiple domains. Second, we present a new paradigm for modeling social-network users and mining time-varying attributes using dynamic bayesian networks (DBNs). We then explore the extent to which such temporal models, combined with collective, semi-supervised training algorithms, can improve the inference results of five behavioral intentions given temporal, real-world social-network data. This work is the first to take a DBN-based approach to the task of private-attribute inference in social networks.

ABSTRACT. The proposed GEP-NB classifier is based on the oversampling technique. It combines two learning methods Gene Expression Programming and Naïve Bayes, which cooperate to produce a final prediction. At the pre-processing stage, a simple mechanism for generating synthetic minority class examples and balancing the training set is used. Next, two genes g1 and g2 are evolved using Gene Expression Programming. They differ by applying in each case a different procedure for selecting synthetic minority class examples. If the class prediction by g1 agrees with the class prediction made by g2, their decision is final. Otherwise, the final predictive decision is taken by the Naïve Bayes classifier. The approach is validated in an extensive computational experiment. Results produced by GEP-NB are compared with the performance of several state-of-the-art classifiers. Comparisons show that GEP-NB offers a competitive performance.

ABSTRACT. Ethereum is a major public blockchain. Besides being the second-largest digital currency by market capitalization for its cryptocurrency, the Ether ETH, it is also the foundation of Web3 and decentralized applications, or DApps, that are fueled by Smart Contracts. Ethereum uses Proof of Work (PoW) consensus algorithm to ensure the integrity of the blockchain and to prevent double spend. PoW requires the participation of miners, who are incentivized to assemble blocks of transactions by being rewarded with cryptocurrency paid by transaction originators and by the blockchain network itself via newly minted ETH. Network fees for transaction submissions are called gas, by analogy to the fuel used by cars, and are negotiable. They are also highly volatile and hence it is critical to predict the direction they are heading into, so that one can time transaction submissions, when feasible. There have been several efforts to predict gas prices, including usage of large Mempools, analysis of committed blocks, and more recent ones using Facebook's Prophet model~\cite{prophet}. Our solution uses DeepAR, a model built on recurrent neural networks (RNN) with the ability leverage hundreds of related time series to produce more accurate predictions. Our implementation uses features extracted from the Ethereum main net as well as off-chain data to achieve accurate predictions.

ABSTRACT. In this talk, we will present a lowest-order staggered discontinuous Galerkin method for the coupled Stokes-Darcy problem. The proposed scheme is locally mass conservative and allows nonmatching grids across the interface. The Beavers-Joseph-Saffman interface conditions are imposed straightforwardly without resorting to the Lagrange multiplier. A new regularity operator is developed to favor the analysis. The stability and the convergence error estimates of the proposed scheme will be rigorously analyzed. In addition, several numerical experiments will be presented to verify the proposed theories.

ABSTRACT. Dimensional reduction strategy is an effective approach to derive reliable conceptual models to describe flow in fractured porous media. The fracture aperture is several orders of magnitude smaller than the characteristic size (e.g., the length of the fracture) of the physical problem. We identify the aperture to length ratio as the small parameter epsilon with the fracture permeability scaled as an exponent of epsilon. Our derivation provides models which exhibit two-scale behaviour.

ABSTRACT. This talk is concerned with the numerical solution of compressible fluid flow in a fractured porous medium. The fracture represents a fast pathway (i.e., with high permeability) and is modeled as a hypersurface embedded in the porous medium. We aim to develop fast-convergent and accurate global-in-time domain decomposition (DD) methods for such a reduced fracture model, in which smaller time step sizes in the fracture can be coupled with larger time step sizes in the subdomains. Using the pressure continuity equation and the tangential PDEs in the fracture-interface as transmission conditions, three different DD formulations are derived; each method leads to a space-time interface problem which is solved iteratively and globally in time. Efficient preconditioners are designed to accelerate the convergence of the iterative methods while preserving the accuracy in time with nonconforming grids. Numerical results for two-dimensional problems with nonimmersed and partially immersed fractures are presented to illustrate the performance of the proposed methods. This is joint work with Yanzhao Cao and Phuoc-Toan Huynh.

ABSTRACT. In this talk, we present numerical approximations of a phase-field model for two-phase ferrofluids, which consists of the Navier-Stokes equations, the Cahn-Hilliard equation, the magnetostatic equations, as well as the magnetic field equation. By combining the projection method for the Navier-Stokes equations and some subtle implicit-explicit treatments for coupled nonlinear terms, we construct a decoupled, linear, fully discrete finite element scheme to solve the highly nonlinear and coupled multi-physics system efficiently. The scheme is provably unconditionally energy stable and leads to a series of decoupled linear equations to solve at each time step. Through numerous numerical examples in simulating benchmark problems such as the Rosensweig instability and droplet deformation, we demonstrate the stability and accuracy of the numerical scheme.

ABSTRACT. Poly lactic-co-glycolic acid (PLGA) is a versatile polymer that can be used to manufacture nano- or micro-scale particles for encapsulating therapeutic agents. Drug release from PLGA particles involves various physical and material properties of PLGA and the encapsulated drug. Diffusion through polymer, convection through pores, osmotic pumping, and degradation have been observed in "in vitro" experiments. However, little progress has been made in mathematical modeling and numerical simulations of this type of drug delivery. In this talk, we present our exploratory work in this approach including validation with experimental data.

ABSTRACT. We introduce Finch, a Julia-based domain specific language (DSL) for solving partial differential equations in a discretization agnostic way, currently including finite element and finite volume methods. A key focus is code generation for various internal or external software targets. Internal targets use a modular set of tools in Julia providing a direct solution within the framework. In contrast, external code generation produces a set of code files to be compiled and run with external libraries or frameworks. Examples include a Matlab target, for smaller problems or prototyping, or C++/MPI based targets for larger problems needing scalability. This allows us to take advantage of their capabilities without needlessly duplicating them, and provides options tailored to the needs of the domain scientist. The modular design of Finch allows ongoing development of these target modules resulting in a more extensible framework and a broader set of applications. The support for multiple discretizations, including finite element and finite volume methods, also contributes to this goal. Another focus of this project is complex systems containing a large set of coupled PDEs that could be challenging to efficiently code and optimize by hand, but that are relatively simple to specify using the DSL. In this paper we present the key features of Finch that set it apart from many other DSL options, and demonstrate the basic usage and current capabilities through examples.

ABSTRACT. The existing mesh refinement approach generates an optimal finite element mesh in an iterative manner. However, this approach requires the user to have prior knowledge of complex PDE problems and error estimation methods. Furthermore, an expensive error estimation process has to be repeatedly performed. The present study was motivated by MeshingNet [1], a pioneering work that performs mesh refinement using neural nets. In MeshingNet, the proposed neural net learns and predicts a target area upper bound, which is the criterion to refine the elements, for a given PDE problem and geometry, and a high-quality non-uniform mesh is generated using this criterion. However, the neural net proposed in MeshingNet has limitations because it assumes that polygons with the same number of boundary vertices are used as inputs as well as outputs. In addition, the PDE error prediction method used is inefficient and lacks robustness. We propose a new mesh refinement framework for solving PDEs based on deep learning. The proposed framework models iterative mesh refinement problems as multi-class classification problems and proposes a new neural net for solving them. Unlike the previous studies, the proposed neural net is flexible in that it can be applied to geometry and boundary conditions having various polygonal shapes. We also propose efficient and robust methods that predict PDE errors using structured grids. The experimental results show that the proposed framework successfully generates high-quality non-uniform meshes, and the solution accuracy of these meshes is considerably improved compared to a uniform mesh with a similar number of elements.

ABSTRACT. This paper is devoted to increasing the computational efficiency of the finite-difference methods for solving the one-way Helmholtz equation in unbounded domains. The higher-order rational approximation of the propagation operator was taken as a basis. Computation of appropriate approximation coefficients and grid sizes is formulated as the problem of minimizing the discrete dispersion relation error. Keeping in mind the complexity of the developed optimization problem, the differential evolution method was used to tackle it. The proposed method does not require manual selection of the artificial parameters of the numerical scheme. The stability of the scheme is provided by an additional constraint of the optimization problem. A comparison with the Padé approximation method and rational interpolation is carried out. The effectiveness of the proposed approach is shown.

ABSTRACT. The root-finding problem is very important in many applications and has become an extensive research field. One of the directions in this field is the use of various iteration schemes. In this paper, we propose a new generalised iteration scheme. The schemes like Mann, Ishikawa, Das-Debata schemes are special cases of the proposed iteration. Moreover, we use the proposed iteration with the PSO-based Newton-like method in two tasks. In the first task, we search for the roots, whereas in the second one for patterns with aesthetic features. The obtained results show that the proposed iteration can decrease the average number of iterations needed to find the roots and that we can generate patterns with potential artistic applications.

ABSTRACT. In the era of data abundance and machine learning technologies, we often encounter difficulties in learning data-driven discovery of hidden physics, that is, learning differential equations/fractional differential equations via data. In \cite{schaeffer2017learningPDE}, Schaeffer proposed a machine learning algorithm to learn the differential equation via data discovery. We extend Schaeffer's work in the case of time fractional differential equations and propose an algorithm to identify the fractional order $\alpha$ and discover the form of $\mathcal{F}$. Furthermore, if we have prior information regarding the set in which parameters belong to have some advantages in terms of time complexity of the algorithm over Schaeffer's work. Finally, we conduct various numerical experiments to verify the method's robustness at different noise levels.

ABSTRACT. The emergence of information theft poses a serious threat to mobile users. Short message service (SMS), as a mainstream communication medium, is usually used by attackers to implement propagation, command and control. The previous detection works are based on the local perspective of terminals, and it is difficult to find all the victims and covert attackers for a theft event. In order to address this problem, we propose DITA-NCG, a method that globally detects information theft attacks based on node communication graph (NCG). The communication behavior of a NCG's node is expressed by both call detail record (CDR) vectors and network flow vectors. Firstly, we use CDR vectors to implement social subgraph division and find suspicious subgraphs with SMS information entropy. Secondly, we use network flow vectors to distinguish information theft attack graphs from suspicious subgraphs, which help us to identify information theft attack. Finally, we evaluate DITA-NCG by using real world network flows and CDRs, and the result shows that DITA-NCG can effectively and globally detect information theft attack in mobile network.

ABSTRACT. Single packet attack, which is initiated by adding attack information to traffic packets, pose a great threat to cybersecurity. Existing detection methods for single packet attack just learn features directly from single packet but ignore the hierarchical relationship of packet resources, which trends to high false positive rate and poor generalization. In this paper, We conduct an extensive measurement study of the realistic traffic and find that the hierarchical relationship of resources is suitable for identifying single packet attacks. Therefore, we propose HNOP, a deep neural network model equipped with the hierarchical relationship, to detect single packet attacks from raw HTTP packets. Firstly, we construct resource node hopping structure based on the ``Referer'' field and the ``URL'' field in HTTP packets. Secondly, hopping features are extracted from the hopping structure of the resource nodes by G_BERT, which are further combined with the lexical features extracted by convolution operation from each node of the structure to form feature vectors. Finally, the extracted features are fed to a classifier, mapping the extracted features to the classification space through a fully connected network, to detect attack traffic. Experiments on the publicly available dataset CICIDS-2017 demonstrate the effectiveness of HNOP with an accuracy of 99.92% and a false positive rate of 0.12%. Furthermore, we perform extensive experiments on dataset IIE_HTTP collected from important service targets at different time. At last, it is verified that the HNOP has the least degraded performance and better generalization compared to the other models.

ABSTRACT. Tor enables end user the desirable cyber anonymity with obfuscation technologies like MEEK. However, it has also manifested itself a wide shield for various illegal hidden services involved cyber criminals, motivating the urgent need of deanonymization technologies. In this paper, we propose a novel communication fingerprint abstracted from key packet sequences, and attempt to efficiently identify end users’ MEEK-based access to Tor hidden services. Specifically, we investigate the communication fingerprint during the early connection stage of MEEK-based Tor rendezvous establishment, and make use of deep neural network to automatically learn and form a key packet sequence. Unlike most of existing approaches that rely on the entire long communication packet sequence, experiments demonstrate that our key packet sequence enabled scheme can significantly reduce both the time and hardware resource consumption for the identification task by 23%-37% and 80%-86%, respectively, while being able to keep a slightly better accuracy.

ABSTRACT. The PageRank algorithm is a widely used linear algebra method with many applications. As graphs with billions or more of nodes become increasingly common, being able to scale this algorithm on modern HPC architectures is of prime importance. While most existing approaches have explored distributed computing to compute an approximation of the PageRank scores, we focus on the numerical computation using the power iteration method. We develop and implement a distributed parellel version of the PageRank. This application is deployed on the Fugaku supercomputer, using up to one thousand compute nodes to assess scalability on random stochastic matrices. These large-scale experiments show that the network-on-chip of the A64FX processor acts as an additional level of computation, in between nodes and cores.

ABSTRACT. We observe and analyze usage of the login nodes of the leadership class Summit supercomputer from the perspective of an ordinary user—not a system administrator—by periodically sampling user activities (job queues, running processes, etc.) for two full years (2020–2021). Our findings unveil key usage patterns that evidence misuse of the system, including gaming the policies, impairing I/O performance, and using login nodes as a sole computing resource. Our analysis highlights observed patterns for the execution of complex computations (workflows), which are key for processing large-scale applications.

ABSTRACT. More and more users are attracted by P2P networks characterized by decentralization, autonomy and anonymity. The management and optimization of P2P networks have become the important research contents. However, most existing methods solve the problem based on the global view of the P2P network which cannot be applied in some cases, e.g. the anti-tracking network. So, the self-evolving of P2P network have become an important research issue. This paper presents a framework for network self-evolving problem based on distributed swarm intelligence, which is achieved by the collaboration of all nodes in a network to evolve their local topologies according to optimal topology model. Each node, as an independent agent, only has the information of its local topology. Through the consensus method, each node searches for an evolving structure of its local topology which is beneficial to all relevant nodes for local topology’s evolving. The self-evolving of each node’s local topology makes the whole topology converge to the optimal topology model. In the experiments, two simulated examples under different network topologies illustrate the feasibility of our approach.

ABSTRACT. In bioinformatics field, the Deep Learning (DL) models offer exceptional results in applications with histological images, scans and tomographies. However, with gene expression data, their performance is limited, further hampered by the complexity of these models that require many instances in the dataset to provide good results. Due to the difficult and expensive collection of medical data, the application of Data Augmentation (DA) techniques is a topic of great importance in the problems currently addressed in bioinformatics. By applying state-of-the-art models based on Conditional Generative Adversary Networks (CGAN), and introducing modifications to the standard method, we investigated the effect of DA to predict the vital status of a patient from RNA-Seq gene expression data. Experimental results on several real-world datasets demonstrate the effectiveness and efficiency of the proposed models. The application of DA methods significantly increase the prediction accuracy, leading by 12% with respect to benchmark datasets and 3.15% with respect to data processed with feature selection. Results obtained with CGAN-based models improve in most cases the SMOTE or noise injection results.

ABSTRACT. Niemann-Pick Class 1 (NPC1) disease is a rare and neurodegenerative disease, and often metabolomics datasets of NPC1 patients are limited in the number of samples and severely imbalanced. In order to improve the predictive capability and identify new biomarkers in an NPC1 disease urinary dataset, data augmentation (DA) techniques based on computational intelligence are employed to create additional synthetic samples. This paper presents DA techniques, based on the addition of noise, on oversampling techniques and using conditional generative adversarial networks, to evaluate their predictive capacities on a set of Nuclear Magnetic Resonance (NMR) profiles of urine samples. Prediction results obtained show increases in sensitivity (30%) and in F1 score (20%). In addition, multivariate data analysis and variable importance in projection scores have been applied. These analyses show the ability of the DA methods to replicate the information of the metabolites and determined that selected metabolites (such as 3-aminoisobutyrate, 3-hidroxivaleric, quinolinate and trimethylamine) may be valuable biomarkers for the diagnosis of NPC1 disease.

ABSTRACT. Readmission rate is an important indicator of the hospital quality of care. With the upsetting increase of readmission rates worldwide, especially in geriatric patients, predicting unplanned readmissions becomes a very im-portant task, that can help to improve the patient’s well-being and reduce healthcare costs. With the aim of reducing hospital readmission more atten-tion is to be paid to the home healthcare services, since home healthcare pa-tients in average have more difficult health conditions. Machine Learning and Artificial intelligence algorithms were used to develop predictive mod-els using MIMIC-IV repository. Developed predictive models account for various patient details, including demographical, administrative, disease re-lated and prescription related data. Categorical features were encoded with a novel customized target encoding approach to improve the model perfor-mance avoiding data leakage and overfitting. This new risk-score based tar-get encoding approach demonstrated similar performance to existing target encoding and Bayesian encoding approaches, with reduced data leakage, when assessed using Gini-importance. Developed models were tested on isolated sets and demonstrated a good discriminative performance, AUC 0.75, TPR 0.69 TNR 0.67 for the best model. These encouraging results, as well as effective feature engineering approach can be used in further studies to develop more reliable 30-day readmission predictive models.

ABSTRACT. Pulmonary Hypertension (PH) is a severe disease characterized by an elevated pulmonary artery pressure. The gold standard for PH diagnosis is measurement of mean Pulmonary Artery Pressure (mPAP) during an invasive Right Heart Catherization. In this paper, we investigate noninvasive approach to PH detection utilizing Magnetic Resonance Imaging, Computer Models and Machine Learning. We show using the ablation study, that physics-informed feature engineering based on models of blood circulation increases the performance of Gradient Boosting Decision Trees-based algorithms for classification of PH and regression of values of mPAP. We compare results of regression (with thresholding of estimated mPAP) and classification and demonstrate that metrics achieved in both experiments are comparable. The predicted mPAP values are more informative to the physicians than the probability of PH returned by classification models. They provide the intuitive explanation of the outcome of the machine learning model (clinicians are accustomed to the mPAP metric, contrary to the PH probability).

ABSTRACT. Post-thrombotic syndrome (PTS) has variable clinical presentation with significant treatment costs and gaps in the evidence-base to support clinical decision making.  The contribution of variations in venous anatomy to the risk of complications following treatment has yet to be characterized in detail.  We report the development of a steady-state, 0D model of venous anatomy of the lower limb and assessments of local sensitivity (10 percent radius variation) and global sensitivity (50 percent radius variation) of the resulting flows to variability in venous anatomy. An analysis of orthogonal sensitivity was also performed.  Local sensitivity analysis was repeated with four degrees of thrombosis in the left common iliac vein. The largest normalised sensitivities were observed in locations associated with the venous return. Both local and global approaches provided similar ranking of input parameters responsible for the variation of flow in a vessel where thrombosis is typically observed.  When a thrombus was included in the model increase in absolute sensitivity was observed in the leg affected by the thrombosis. These results can be used to inform model reduction strategies and to target clinical data collection.

ABSTRACT. We improve the utility of Risk-calibrated Supersparse Linear Integer Model (RiskSLIM). It is a scoring systems which is an interpretable machine learning classification model optimized for performance. Scoring systems are commonly used in healthcare and justice. We implement feature discretization (FD) in the hyperparameter optimization process in order to improve classification performance and refer to the new approach as FD-RiskSLIM. We test the approach using two medical applications. We compare the results of FD-RiskSLIM, RiskSLIM, and other machine learning (ML) models. We demonstrate that scoring models based on RiskSLIM, in addition to being interpretable, perform at least on par with the state-of-the-art ML models such as Gradient Boosting in terms of classification metrics. We show superiority of FD-RiskSLIM over RiskSLIM.

ABSTRACT. This paper describes the automated production data acquisition and integration process in the architectural pattern Tweeting Factory. This concept allows the use of existing production equipment with PLCs and the use of industrial IoT prepared for Industry 4.0. The main purpose of the work is to propose an event-driven system architecture and to prove its correctness and efficiency. The proposed architecture is able to perform transformation operations on the collected data. The simulation tests were carried out using real data from the factory shop-floor, services prepared for production monitoring, allowing the calculation of KPIs. The correctness of the solution is confirmed on 20 production units by comparing its results with the blackboard architecture using SQL queries. Finally, the response time for calculating ISO 22400 performance indicators is examined and it was verified that the presented solution can be considered as a real-time system.

ABSTRACT. A classical NP-hard problem is the Minimum Edge Clique Cover (minECC) problem, which is concerned with covering edges of a network (graph) with the minimum number of cliques. There are many real-life applications of this problem, such as in food science, computational biology, efficient representation of pairwise information, and so on. Borrowing ideas from [7], we propose using a compact representation, the intersection representation, of network data and design an efficient and scalable algorithm for minECC. Edges are considered for inclusion in cliques in degree-based orders during the clique construction step. The intersection representation of the input graph enabled efficient computer implementation of the algorithm by utilizing an existing sparse matrix package [10]. We present results from numerical experiments on a representative set of real-world and synthetically constructed benchmark graph instances. Our algorithm significantly outperforms the current state-of-the-art heuristic algorithm of [4] in terms of the quality of the edge clique covers returned and running time performance on the benchmark test instances. On some of the largest graph instances, existing heuristics failed to terminate while our algorithm could finish the computation within a reasonable amount of time.

ABSTRACT. Science collaborations such as ATLAS at the high-energy particle accelerator at CERN use a computer grid to run expensive computational tasks on huge, distributed data sets.

Dealing with big data on a grid demands work load management and data allocation to maintain a continuous workflow. Data allocation in a computer grid necessitates some data placement policy that is conditioned on the resources of the system and the usage of data.

In part, automatic and manual data policies shall achieve short time-to-result. There are efforts to improve data policies. Data allocation is important to cope with the increasing amount of data and processing on different data centers. A data allocation or data placement policy decides at which location, sub sets of data are to be placed.

In this paper, a novel approach cope with the problem of the bottleneck related to wide area file transfers between data centers and of large distributed data sets with high dimensionality. The model estimates similar data with a neural network on sparse and uncertain observations, and then proceeds with the allocation process. The allocation process comprises evolutionary data allocation for finding near-optimal solutions, and improves over 5 on network transfers for the given data centers.

ABSTRACT. Triangles are an essential part of network analysis, representing metrics such as transitivity and clustering coefficient. Using the correspondence between sparse adjacency matrices and graphs, linear algebraic methods have been developed for triangle counting and enumeration, where the main computational kernel is sparse matrix-matrix multiplication. In this paper, we use an intersection representation of graph data implemented as a sparse matrix, and engineer an algorithm to compute the “k-count” distribution of the triangles of the graph. The main computational task of computing sparse matrix-vector products is carefully crafted by employing compressed vectors as accumulators. Our method avoids redundant work by counting and enumerating each triangle exactly once. We present results from extensive computational experiments on large-scale real-world and synthetic graph instances that demonstrate good scalability of our method. In terms of run-time performance, our algorithm has been found to be orders of magnitude faster than the reference implementations of the miniTri data analytics application.

ABSTRACT. Changes in the demographic structure of the population have imposed alter-ations in the pension systems. In many countries, including Poland, the amount of retirement benefits is highly dependent on life expectancy, which in the case of increases in longevity leads to a decrease in accrued benefits. A dynamic Monte Carlo simulation model was developed to investigate the fi-nancial implications of the aging problem in connection with the previously unexpected demographic changes caused by the Covid-19 pandemic on fu-ture pension payments. The model uses data from Polish statistical data-bases. The study distinguishes different life cycle profiles, i.e. women and men with average and minimum wage earnings. Simulation experiments are conducted in two variants. The first variant takes into account the currently registered shortening of life expectancy, while the second variant assumes that life expectancy is continuously lengthening, as it was observed until the outbreak of the Covid-19 epidemic. The simulation results show that the Covid-19 pandemic has a beneficial effect for future retirees, which is re-flected in the expected higher replacement rates at retirement.

ABSTRACT. In nature, deoxyribonucleic acid (DNA) contains the genetic instructions used in the development and functioning of all known living organisms. The idea behind our vision is to develop an artificial system, an architecture that would support discovering, adding, storing, improving and sharing information and knowledge among agents and organizations through experience. We propose a novel Knowledge Representation (KR) approach in which experiential knowledge is represented by Set of Experience (SOE), and is carried into the future by Decisional DNA (DDNA). This research has enormous and exciting potential of opening entirely new and so far not conceptually conceived horizons in developing and using collective intelligence to find solutions to different problems. DDNA has been applied in different domains . DDNA has been applied in different domains augmenting capabilities, facilitating smart knowledge management, and engineering processes inside decision-making Decisional DNA has been applied in different domains augmenting capabilities, facilitating smart knowledge management, and engineering processes inside decision-making procedures. This paper introduces the newest and probably the furthermost large-scale DDNA extension which  is the Idream.Technology platform (https://www.idream.technology). We are developing social digital platform using collective experience. This DDNA-based application, which commences in the second half of  2022, projects personalized road-maps  to achieve purposes, goals, and aims taking into account individuals’ personalities and circumstances. Specific areas of human activities covered within the platform include travel, education, acquisition, and well-being. Idream.Technology is a start-up hi-tech company which was launched in May of 2021 and is ready to offer its smart enhancement of dedicated coaching services.  It provides DDNA technology based tools to enhance peoples’ sustainable development.

ABSTRACT. We develop an empirically grounded agent-based model to explore the purchasing decisions of mutually interacting agents (consumers) between three types of alternative fuel vehicles. We calibrate the model with recently published empirical data on consumer preferences towards such vehicles. Furthermore, running the Monte Carlo simulations, we show possible scenarios for the development of the alternative fuel vehicle market depending on the marketing strategies employed.

ABSTRACT. In this paper, we propose an additional step in tuning information retrieval systems after dataset experiments and before testing them with real users. For this purpose, we add a simulation of the group of users as a social collective model and run the information retrieval system on it. Here, we take first steps in that direction, by focusing on a subsystem of a recommender system and simulating what type of social collective it is most effective with. We present details on the social collective model and the information retrieval subsystem model, as well as how to put them together in one simulation. We run several experiments and present some initial findings. In our opinion, this overall approach could be used to greatly enhance further tests with real users.

ABSTRACT. Sentiments expressed in opinions on social networks have played an increasingly significant impact in solving various social problems. Improving the effectiveness of sentiment analysis methods on social networks is still of interest to several scientists. A notable and robust development direction is sentiment analysis methods based on graph convolutional networks (GCNs). This paper introduces a model called contextual-based GCN for sentence-level sentiment analysis by considering the following steps: (i) Sentences are converted into contextualized word representation vectors based on the combination of the bidirectional encoder representations from the transformer model and bidirectional long short-term memory. (ii) The contextualized word representations are used to construct a sentence graph as a feature of nodes. (iii) A GCN model with two convolutional layers was used to learn the structure-aware node representations on the sentence graph. (iv) The softmax classifier was used for the sentence-level sentiment analysis. Experimental results on benchmark datasets showed that, unlike other methods, the proposed method can extract more context information from the opinions to obtain a better representation of the graph structure and learn better structure-aware nodes represented on the graph. The proposed method has improved the performance in terms of accuracy of the conventional methods from 2.2 to 3.2 percentage points.

ABSTRACT. To push out a blood flow to an aorta, a left ventricle repeats expansion and contraction motion. For more efficient pumping of the blood, it is known that the left ventricle also has twisted motion. In this paper, the influence of the twisted motion for a blood flow to an aorta was investigated. In particular, the relationship between the origin of cardiovascular disease and wall shear stress has been pointed out in the aorta region. Estimating the difference of the wall shear stress depend on the presence or absence of the twisted motion, the blood flow simulation was conducted. To express its complicated shape and the motion, the unstructured moving grid finite volume method was adopted. In this method, the control volume is defined for a space time unified domain. Not only a physical conservation law but also a geometric conservation law is satisfied in this approach. Then high accurate computation is conducted under the method. From the computation results, a remarkable difference of complicated vortex structures generated in the left ventricle was found as the influence of the left ventricular twisted motion. The vortex structures affected the blood flow leading into the aorta with the result that they generated a clear difference of the wall shear stress. The region where the difference occurred is aortic arch, then it corresponded with a favorite site of arteriosclerosis. Thus, the result showed the possibility that the simulation with the left ventricular twisted motion would be useful to specify causes of heart diseases.

ABSTRACT. In this study, we conducted a computational fluid dynamics analysis to estimate the trajectory of the virus-laden droplets. As numerical models, two human body models with airways were prepared. These models are represented by unstructured grids. Having calculated the unsteady airflow in the room, we simulated the trajectory of droplets emitted by the human speaking. In addition, inhaling the droplets into the lung of the conversation partner was simulated. The number of the droplets adhered to the respiratory lining of the partner was counted separately on the nasal cavity, oral cavity, trachea, bronchi, and bronchial inlet surface. The diameters of the droplets were also investigated in the same manner. It was noticeable that more than 80% of the droplets inhaled by the conversation partner adhered to the bronchial inlet surface. Also, the conversation partner did not inhale droplets larger than 35μm in diameter. It was found that when the distance between two people was 0.75m, more droplets adhered to the partner’s torso.

ABSTRACT. Hydrogen is an ideal clean energy and the transportation of hydrogen is a significant link in hydrogen energy industry. It is an economical and efficient way to realize the long-distance transportation of hydrogen and urban distribution of hydrogen-enriched natural gas through polyethylene pipelines. However, the hydrogen can permeate polyethylene pipelines because of the small size of hydrogen molecules, resulting in explosion risk and loss of hy-drogen. In this paper, the solubility, diffusion and permeation characteristics of hydrogen in amorphous polyethylene at temperature from 270 to 310 K and pressure from 0.109 to 4.0 MPa are studied by the molecular dynamics (MD) simulation, and the permeation of hydrogen in polyethylene pipelines is analyzed and evaluated. Results illustrate that both the diffusion coefficient and permeability coefficient increase with the rise of temperature, while the solubility shows an opposite trend. Due to the small range of pressure, the variation of diffusion and permeability coefficients is ignorable with the increase of pressure, indicating the pressure change has slight effects on diffusion and permeability characteristics. It is also found that the change of permeability coefficient is mainly caused by the variation of diffusion coefficient, that is, the variation of diffusion coefficient is a dominant factor influencing the permeation of hydrogen in polyethylene pipelines.

ABSTRACT. This study achieves modelling two helicopters via computational fluid dynamics (CFD) and simulating the flow field that develops due to a near miss. The rotation of the main rotor and the translational movement of the helicopter are modelled in this study, and the long trajectory of the moving helicopter is realised in the simulation. Moreover, the interaction of flows around the two moving helicopters is also achieved by introducing the communication between multiple moving computational domains. Firstly, the validation test is conducted using a helicopter model with a rotating main rotor, where the results produced by our in-house code are compared with those computed by another CFD solver, FaSTAR-Move. This test verifies that the communication between the overlapping grids is reliably achieved in our simulation. In the simulation of nearly missing helicopters, two near-miss cases are computationally demonstrated, where the complex flow field which develops around the two helicopters is captured, and the disturbance in aerodynamic and moment coefficients exerted on the helicopters are observed. These results confirm the capability of this CFD approach for realising near-miss events on a computer.