Computational Science – ICCS 2021

In this study, we analyze and identify a proper scale value when presenting real world space and everyday objects in immerse VR. We verify the impact of usage of reference points in the form of common objects known to the user such as windows, doors and furniture in the sense of scale in VR. We also analyze user behavior (position, rotation, movement, area of interest and such) in the scale setting task. Finally, we propose optimal scale values for single objects presentation, architectural space with many points of references and a large scale space with less to no points of reference. The experiments were conducted on two groups: the Experts (architects) and Non-experts (common users) to verify the translation of real-world object size analysis skills into the same capacity in the virtual world. Confirmation of the significance of the pre-immersion in VR for a sense of scale accuracy is also described.

Many concepts behind Capsule Networks cannot be proved due to limited research, performed so far. In the paper, we compare the CapsNet architecture with the most common implementations of convolutional networks (CNNs) for image classification. We also introduced Convolutional CapsNet - a network that mimics the original CapsNet architecture but remains a pure CNN - and compare it against CapsNet. The networks are tested using popular benchmark image data sets and additional test sets, specifically generated for the task. We show that for a group of data sets, usage of CapsNet-specific elements influences the network performance. Moreover, we indicate that the use of Capsule Network and CNN may be highly dependent on the particular data set in image classification.

Since diverse and complex emotions need to be expressed by different facial deformation and appearances, facial animation has become a serious and on-going challenge for computer animation industry. Face reconstruction techniques based on 3D morphable face model and deep learning provide one effective solution to reuse existing databases and create believable animation of new characters from images or videos in seconds, which greatly reduce heavy manual operations and a lot of time. In this paper, we review the databases and state-of-the-art methods of 3D face reconstruction from a single RGB image. First, we classify 3D reconstruction methods into three categories and review each of them. These three categories are: Shape-from-Shading (SFS), 3D Morphable Face Model (3DMM), and Deep Learning (DL) based 3D face reconstruction. Next, we introduce existing 2D and 3D facial databases. After that, we review 10 methods of deep learning-based 3D face reconstruction and evaluate four representative ones among them. Finally, we draw conclusions of this paper and discuss future research directions.

Time varying meshes are more popular than ever as a representation of deforming shapes, in particular for their versatility and inherent ability to capture both true and spurious topology changes. In contrast with dynamic meshes, however, they do not capture the temporal correspondence, which (among other problems) leads to very high storage and processing costs. Unfortunately, establishing temporal correspondence of surfaces is difficult, because it is generally not bijective: even when the full visible surface is captured in each frame, some parts of the surface may be missing in some frames due to self-contact. We observe that, in contrast with the inherent absence of bijectivity in surface correspondence, volume correspondence is bijective in a wide class of possible input data. We demonstrate that using a proper intitialization and objective function, it is possible to track the volume, even when considering only a pair of subsequent frames at the time. Currently, the process is rather slow, but the results are promising and may lead to a new level of understanding and new algorithms for processing of time varying meshes, including compression, editing, texturing and others.

Since the development of the touch screen technology makes sketches simple to draw and obtain, sketch-based 3D shape retrieval has received increasing attention in the community of computer vision and graphics in recent years. The main challenge is the big domain discrepancy between 2D sketches and 3D shapes. Most existing works tried to simultaneously map sketches and 3D shapes into a joint feature embedding space, which has a low efficiency and high computational cost. In this paper, we propose a novel semantic similarity metric learning method based on a teacher-student strategy for sketch-based 3D shape retrieval. We first extract the pre-learned semantic features of 3D shapes from the teacher network and then use them to guide the feature learning of 2D sketches in the student network. The experiment results show that our method has a better retrieval performance.

Charts or scientific plots are widely used visualizations for efficient knowledge dissemination from datasets. Nowadays, these charts are predominantly available in image format in print media, the internet, and research publications. There are various scenarios where these images are to be interpreted in the absence of datasets that were originally used to generate the charts. This leads to a pertinent need for automating data extraction from an available chart image. We narrow down our scope to scatter plots and propose a semi-automated algorithm, ScatterPlotAnalyzer, for data extraction from chart images. Our algorithm is designed around the use of second-order tensor fields to model the chart image. ScatterPlotAnalyzer integrates the following tasks in sequence: chart type classification, image annotation, object detection, text detection and recognition, data transformation, text summarization, and optionally, chart redesign. The novelty of our algorithm is in analyzing both simple and multi-class scatter plots. Our results show that our algorithm can effectively extract data from images of different resolutions. We also discuss specific test cases where ScatterPlotAnalyzer fails.

In this day and age, Electroencephalography-based methods for Automated Affect Recognition are becoming more and more popular. Owing to the vast amount of information gathered in EEG signals, such methods provide satisfying results in terms of Affective Computing. In this paper, we replicated and improved the CNN-based method proposed by Li et al. [11]. We tested our model using a Dataset for Emotion Analysis using EEG, Physiological and Video Signals (DEAP) [9]. Performed changes in the data preprocessing and in the model architecture led to an increase in accuracy – 74.37% for valence, 73.74% for arousal.

Deep neural networks are the frontier in object detection, a key modern computing task. The dominant methods involve two-stage deep networks that heavily rely on features extracted by the backbone in the first stage. In this study, we propose an improved model, ResNeXt101S, to improve feature quality for layers that might be too deep. It introduces splits in middle layers for feature extraction and a deep feature pyramid network (DFPN) for feature aggregation. This backbone is neither much larger than the leading model ResNeXt nor increasing computational complexity distinctly. It is applicable to a range of different image resolutions. The evaluation of customized benchmark datasets using various image resolutions shows that the improvement is effective and consistent. In addition, the study shows input resolution does impact detection performance. In short, our proposed backbone can achieve better accuracy under different resolutions comparing to state-of-the-art models.

This paper presents a complete solution for procedural creation of new levels, implemented in an existing puzzle video game. It explains the development, going through an adaptation to the genre of game of the approach to puzzle generation and talking in detail about various difficulty metrics used to calculate the resulting grade. Final part of the research presents the results of grading a set of hand-crafted levels to demonstrate the viability of this method, and later presents the range of scores for grading generated puzzles using different settings. In conclusion, the paper manages to deliver an effective system for assisting a designer with prototyping new puzzles for the game, while leaving room for future performance improvements.

Euler Lagrange Skeletal Animation (ELSA) is the novel and fast model for skeletal animation, based on the Euler Lagrange equations of motion and configuration and phase space notion. Single joint’s animation is an integral curve in the vector field generated by those PDEs. Considering the point in the phase space belonging to the animation at current time, by adding the vector pinned to this point and multiplied by the elapsed time, one can designate the new point in the phase space. It defines the state, especially the position (or rotation) of the joint after this time elapses. Starting at time 0 and repeating this procedure N times, there is obtained the approximation, and if the $$N\rightarrow \infty $$ N → ∞ the integral curve itself. Applying above, to all joint in the skeletal model constitutes ELSA.

Cross-section curves play an important role in many fields. Analytically representing cross-section curves can greatly reduce design variables and related storage costs and facilitate other applications. In this paper, we propose composite generalized elliptic curves to approximate open and closed cross-section curves, present their mathematical expressions, and derive the mathematical equations of surface reconstruction from composite generalized elliptic curves. The examples given in this paper demonstrate the effectiveness and high accuracy of the proposed method. Due to the analytical nature of composite generalized elliptic curves and the surfaces reconstructed from them, the proposed method can reduce design variables and storage requirements and facilitate other applications such as level of detail.

In the paper we address a problem of estimating a physical state of an observed person by means of analysing facial portrait captured in thermal spectrum. The algorithm consists of facial regions detection combined with tracking and individual features classification. We focus on eyes and mouth state estimation. The detectors are based on Haar-like features and AdaBoost, previously applied to visible-band images. Returned face region is subject to eyes and mouth detection. Further, extracted regions are filtered using Gabor filter bank and the resultant features are classified. Finally, classifiers’ responses are integrated and the decision about driver’s physical state is taken. By using thermal image we are able to capture eyes and mouth states in very adverse lighting conditions, in contrast to the visible-light approaches. Experiments performed on manually annotated video sequences have shown that the proposed approach is accurate and can be a part of current Advanced Driver Assistant Systems.

This work contains a solution improvement for Smart Events, which are one of the ways to guide the behavior of NPCs in computer games. The improvement consists of three aspects: introducing the possibility of group actions by agents, i.e. cooperation between them, extending the SE with the possibility of handling ordinary events not only emergency, and introducing the possibility of taking random (but predetermined) actions as part of participation in the event.In addition, two event scenarios were presented that allowed the Smart Events operation to be examined. The study consists of comparing the performance of the SE with another well-known algorithm (FSM) and of comparing different runs of the same event determined by the improved algorithm.Comparing the performance required proposing measures that would allow for the presentation of quantitative differences between the runs of different algorithms or the same algorithm in different runs. Three were proposed: time needed by the AI subsystem in one simulation frame, the number of decisions in the frame, and the number of frames per second of simulation.

Current approaches to visual place recognition for loop closure do not provide information about confidence of decisions. In this work we present an algorithm for place recognition on the basis of graph-based decisions on deep embeddings and blur detections. The graph constructed in advance permits together with information about the room category an inference on usefulness of place recognition, and in particular, it enables the evaluation the confidence of final decision. We demonstrate experimentally that thanks to proposed blur detection the accuracy of scene recognition is much higher. We evaluate performance of place recognition on the basis of manually selected places for recognition with corresponding sets of relevant and irrelevant images. The algorithm has been evaluated on large dataset for visual place recognition that contains both images with severe (unknown) blurs and sharp images. Images with 6-DOF viewpoint variations were recorded using a humanoid robot.

In this paper, we present an end-to-end application to perform automatic multiple player detection, unsupervised labelling, and a semi-automatic approach to finding homographies. We incorporate dense optical flow for modelling camera movement and user-assisted calibration on automatically chosen key-frames. Players detection is performed with a pre-trained YOLOv3 detector and player labelling is done using features in HSV colorspace. The experimental results demonstrate that our method is reliable with generating heatmaps from players’ positions in case of moderate camera movement. Major limitations of proposed method are the necessity of manual calibration of characteristic frames, inaccuracy with fast camera movements, and small tolerance of vertical camera movement.

Partial differential equation (PDE) based geometric modelling has a number of advantages such as fewer design variables, avoidance of stitching adjacent patches together to achieve required continuities, and physics-based nature. Although a lot of papers have investigated PDE-based shape creation, shape manipulation, surface blending and volume blending as well as surface reconstruction using implicit PDE surfaces, there is little work of investigating PDE-based shape reconstruction using explicit PDE surfaces, specially satisfying the constraints on four boundaries of a PDE surface patch. In this paper, we propose a new method of using an accurate closed form solution to a fourth-order partial differential equation to reconstruct 3D surfaces from point clouds. It includes selecting a fourth-order partial differential equation, obtaining the closed form solutions of the equation, investigating the errors of using one of the obtained closed form solutions to reconstruct PDE surfaces from differential number of 3D points.

Missing values are ubiquitous in many real-world datasets. In scenarios where a dataset is not very large, addressing its missing values by utilizing appropriate data imputation methods benefits analysis significantly. In this paper, we leveraged and evaluated a new imputation approach called k-Nearest Neighbour with Most Significant Features and incomplete cases (KNNI $$_\mathrm{MSF}$$ MSF ) to impute missing values in a healthcare dataset. This algorithm leverages k-Nearest Neighbour (kNN) and ReliefF feature selection techniques to address incomplete cases in the dataset. The merit of imputation is measured by comparing the classification performance of data models trained with the dataset with imputation and without imputation. We used a real-world dataset, “very low birth weight infants”, to predict the survival outcome of infants with low birth weights. Five different classifiers were used in the experiments. The comparison of multiple performance metrics shows that classifiers built on imputed dataset produce much better outcomes. KNNI $$_\mathrm{MSF}$$ MSF outperformed in general than the k-Nearest Neighbour Imputation using the Random Forest feature weights (KNNI $$_\mathrm{RF}$$ RF ) algorithm with respect to the balanced accuracy and specificity.

This paper shows how a gradient-free optimization method is used to improve the prediction capabilities of wildfire progression by estimating the wind conditions driving a FARSITE wildfire model. To characterize the performance of the prediction of the perimeter as a function of the wind conditions, an uncertainty weighting is applied to each vertex of the measured fire perimeter and a weighted least-squares error is computed between the predicted and measured fire perimeter. In addition, interpolation of the measured fire perimeter and its uncertainty is adopted to match the number of vertices on the predicted and measured fire perimeter. The gradient-free optimization based on iterative refined gridding provides robustness to intermittent erroneous results produced by FARSITE and quickly find optimal wind conditions by paralleling the wildfire model calculations. Results on wind condition estimation are illustrated on two historical wildfire events: the 2019 Maria fire that burned south of the community of Santa Paula in the area of Somis, CA, and the 2019 Cave fire that started in the Santa Ynez Mountains of Santa Barbara County.

Handling big data has largely been a major bottleneck in traditional statistical models. Consequently, when accurate point prediction is the primary target, machine learning models are often preferred over their statistical counterparts for bigger problems. But full probabilistic statistical models often outperform other models in quantifying uncertainties associated with model predictions. We develop a data-driven statistical modeling framework that combines the uncertainties from an ensemble of statistical models learned on smaller subsets of data carefully chosen to account for imbalances in the input space. We demonstrate this method on a photometric redshift estimation problem in cosmology, which seeks to infer a distribution of the redshift—the stretching effect in observing the light of far-away galaxies—given multivariate color information observed for an object in the sky. Our proposed method performs balanced partitioning, graph-based data subsampling across the partitions, and training of an ensemble of Gaussian process models.

The problem of limited labelled data availability causes under-fitting, which negatively affects the development of accurate time series based prediction models. Two-hybrid deep neural network architectures, namely the CNN-BiLSTM and the Conv-BiLSTM, are proposed for time series based transductive transfer learning and compared to the baseline CNN model. The automatic feature extraction abilities of the encoder CNN module combined with the superior recall of both short and long term sequences by the decoder LSTM module have shown to be advantageous in transfer learning tasks. The extra ability to process in both forward and backward directions by the proposed models shows promising results to aiding transfer learning. The most consistent transfer learning strategy involved freezing both the CNN and BiLSTM modules while retraining only the fully connected layers. These proposed hybrid transfer learning models were compared to the baseline CNN transfer learning model and newly created hybrid models using the $$R^2$$ R 2 , MAE and RMSE metrics. Three electrical vehicle data-sets were used to test the proposed transfer frameworks. The results favour the hybrid architectures for better transfer learning abilities relative to utilising the baseline CNN transfer learning model. This study offers guidance to enhance time series-based transfer learning by using available data sources.

Buildings consume about 40% of the world's energy use. Energy efficiency in buildings is an increasing concern for the building owners. A reliable energy use prediction model is crucial for decision-makers. This study proposed a hybrid machine learning model for predicting one-day-ahead time-series electricity use data in buildings. The proposed SAMFOR model combined support vector regression (SVR) and firefly algorithm (FA) with conventional time-series seasonal autoregressive integrated moving average (SARIMA) forecasting model. Large datasets of electricity use in office buildings in Vietnam were used to develop the forecasting model. Results show that the proposed SAMFOR model was more effective than the baselines machine learning models. The proposed model has the lowest errors, which yielded 0.90 kWh in RMSE, 0.96 kWh in MAE, 9.04% in MAPE, 0.904 in R in the test phase. The prediction results provide building managers with useful information to enhance energy-saving solutions.

Road closures due to adverse and severe weather continue to affect Wyoming due to hazardous driving conditions and temporarily suspending interstate commerce. The mountain ranges and elevation in Wyoming makes generating accurate predictions challenging, both from a meteorological and machine learning stand point. In a continuation of prior research, we investigate the 80 km stretch of Interstate-80 between Laramie and Cheyenne using autonomous machine learning to create an improved model that yields a 10% increase in closure prediction accuracy. We explore both serial and parallel implementations run on a supercomputer. We apply auto-sklearn, a popular and well documented autonomous machine learning toolkit, to generate a model utilizing ensemble learning. In the previous study, we applied a linear support vector machine with ensemble learning. We will compare our new found results to previous results.

In this paper, a Conv-BiLSTM hybrid architecture is proposed to improve building energy consumption reconstruction of a new multi-functional building type. Experiments indicate that using the proposed hybrid architecture results in improved prediction accuracy for two case multi-functional buildings in ultra-short-term to short term energy use modelling, with $$R^2$$ R 2 score ranging between 0.81 to 0.94. The proposed model architecture comprising the CNN, dropout, bidirectional and dense layer modules superseded the performance of the commonly used baseline deep learning models tested in the investigation, demonstrating the effectiveness of the proposed architectural structure. The proposed model is satisfactorily applicable to modelling multi-functional building energy consumption.

This paper aims to improve short-term forecasting of clouds to accelerate the usability of solar energy. It compares the Convolutional Gated Recurrent Unit (ConvGRU) model to an optical flow baseline and the Numerical Weather Prediction (NWP) Weather Research and Forecast (WRF) model. The models are evaluated over 75 days in the summer of 2019 for an area covering the Netherlands, and it is studied under what circumstance the models perform best. The ConvGRU model proved to outperform both extrapolation-based methods and an operational NWP system in the precipitation domain. For our study, the model trains on sequences containing irradiance data from the Meteosat Second Generation Cloud Physical Properties (MSG-CPP) dataset. Additionally, we design an extension to the model, enabling the model also to exploit geographical data. The experimental results show that the ConvGRU outperforms the other methods in all weather conditions and improves the optical flow benchmark by $$9\%$$ 9 % in terms of Mean Absolute Error (MAE). However, the ConvGRU prediction samples demonstrate that the model suffers from a blurry image problem, which causes cloud structures to smooth out over time. The optical flow model is better at representing cloud fields throughout the forecast. The WRF model performs best on clear days in terms of the Structural Similarity Index Metric (SSIM) but suffers from the simulation’s short-range.

We propose the Automatic-differentiated Physics-Informed Echo State Network (API-ESN). The network is constrained by the physical equations through the reservoir’s exact time-derivative, which is computed by automatic differentiation. As compared to the original Physics-Informed Echo State Network, the accuracy of the time-derivative is increased by up to seven orders of magnitude. This increased accuracy is key in chaotic dynamical systems, where errors grow exponentially in time. The network is showcased in the reconstruction of unmeasured (hidden) states of a chaotic system. The API-ESN eliminates a source of error, which is present in existing physics-informed echo state networks, in the computation of the time-derivative. This opens up new possibilities for an accurate reconstruction of chaotic dynamical states.

We discuss the application of a supervised machine learning method, random forest algorithm (RF), to perform parameter space exploration and sensitivity analysis on ordinary differential equation models. Decision trees can provide complex decision boundaries and can help visualize decision rules in an easily digested format that can aid in understanding the predictive structure of a dynamic model and the relationship between input parameters and model output. We study a simplified process for model parameter tuning and sensitivity analysis that can be used in the early stages of model development.

We present an Auto-Encoded Reservoir-Computing (AE-RC) approach to learn the dynamics of a 2D turbulent flow. The AE-RC consists of an Autoencoder, which discovers an efficient manifold representation of the flow state, and an Echo State Network, which learns the time evolution of the flow in the manifold. The AE-RC is able to both learn the time-accurate dynamics of the flow and predict its first-order statistical moments. The AE-RC approach opens up new possibilities for the spatio-temporal prediction of turbulence with machine learning.

This paper proposes a new decomposition technique for the general class of Non-linear Finite Impulse Response (NFIR) systems. Based on the estimates of projection operators, we construct a set of coefficients, sensitive to the separated internal system components with short-term memory, both linear and nonlinear. The proposed technique allows for the internal structure inference in the presence of unknown additive disturbance on the system output and for a class of arbitrary but bounded nonlinear characteristics.The results of numerical experiments, shown and discussed in the paper, indicate applicability of the method for different types of nonlinear characteristics in the system.

The focus of this study is to simulate realistic fluid flow, through Machine Learning techniques that could be utilised in real-time forecasting of urban air pollution. We propose a novel Latent GAN architecture which looks at combining an AutoEncoder with a Generative Adversarial Network to predict fluid flow at the proceeding timestep of a given input, whilst keeping computational costs low. This architecture is applied to tracer flows and velocity fields around an urban city. We present a pair of AutoEncoders capable of dimensionality reduction of 3 orders of magnitude. Further, we present a pair of Generator models capable of performing real-time forecasting of tracer flows and velocity fields. We demonstrate that the models, as well as the latent spaces generated, learn and retain meaningful physical features of the domain. Despite the domain of this project being that of computational fluid dynamics, the Latent GAN architecture is designed to be generalisable such that it can be applied to other dynamical systems.

Data Assimilation (DA) is a Bayesian inference that combines the state of a dynamical system with real data collected by instruments at a given time. The goal of DA is to improve the accuracy of the dynamic system making its result as real as possible. One of the most popular technique for DA is the Kalman Filter (KF). When the dynamic system refers to a real world application, the representation of the state of a physical system usually leads to a big data problem. For these problems, KF results computationally too expensive and mandates to use of reduced order modeling techniques. In this paper we proposed a new methodology we called Latent Assimilation (LA). It consists in performing the KF in the latent space obtained by an Autoencoder with non-linear encoder functions and non-linear decoder functions. In the latent space, the dynamic system is represented by a surrogate model built by a Recurrent Neural Network. In particular, an Long Short Term Memory (LSTM) network is used to train a function which emulates the dynamic system in the latent space. The data from the dynamic model and the real data coming from the instruments are both processed through the Autoencoder. We apply the methodology to a real test case and we show that the LA has a good performance both in accuracy and in efficiency.

Understanding the community structure of countries in the international food network is of great importance for policymakers. Indeed, clusters might be the key for the understanding of the geopolitical and economic interconnectedness between countries. Their detection and analysis might lead to a bona fide evaluation of the impact of spillover effects between countries in situations of distress. In this paper, we introduce a clustering methodology that we name Higher-order Hierarchical Spectral Clustering (HHSC), which combines a higher-order tensor factorization and a hierarchical clustering algorithm. We apply this methodology to a multidimensional system of countries and products involved in the import-export trade network (FAO dataset). We find a structural proxy of countries interconnectedness that is not only valid for a specific product but for the whole trade system. We retrieve clusters that are linked to economic activity and geographical proximity.

The computational simulation of physical phenomena is a highly complex and expensive process. Traditional simulation models, based on equations describing the behavior of the system, do not allow generating data in sufficient quantity and speed to predict its evolution and make decisions accordingly automatically. These features are particularly relevant in building energy simulations. In this work, we introduce the idea of deep data-driven simulation models (D3S), a novel approach in terms of the combination of models. A D3S is capable of emulating the behavior of a system in a similar way to simulators based on physical principles but requiring less effort in its construction—it is learned automatically from historical data—and less time to run—no need to solve complex equations.

The parameters of a level-set flame model are inferred using an ensemble of heteroscedastic Bayesian neural networks (BayNNEs). The neural networks are trained on a library of 1.7 million observations of 8500 simulations of the flame edge, obtained using the model with known parameters. The ensemble produces samples from the posterior probability distribution of the parameters, conditioned on the observations, as well as estimates of the uncertainties in the parameters. The predicted parameters and uncertainties are compared to those inferred using an ensemble Kalman filter. The expected parameter values inferred with the BayNNE method, once trained, match those inferred with the Kalman filter but require less than one millionth of the time and computational cost of the Kalman filter. This method enables a physics-based model to be tuned from experimental images in real time.

In recent years, Machine Learning (ML) algorithms have proved to be very helpful in several research fields, such as engineering, health-science, physics etc. Among these fields, Astrophysics also started to develop a stronger need of ML techniques for the management of big-data collected by ongoing and future all-sky surveys (e.g. Gaia, LAMOST, LSST etc.). NASA’s Transiting Exoplanet Survey Satellite (TESS) is a space-based all-sky time-domain survey searching for planets outside of the solar system, by means of transit method. During its first two years of operations, TESS collected hundreds of terabytes of photometric observations at a two minutes cadence. ML approaches allow to perform a fast planet candidates recognition into TESS light curves, but they require assimilated data. Therefore, different pre-processing operations need to be performed on the light curves. In particular, cleaning the data from inconsistent values is a critical initial step, but because of the large amount of TESS light curves, this process requires a long execution time. In this context, High-Performance computing techniques allow to significantly accelerate the procedure, thus dramatically improving the efficiency of the outliers rejection. Here, we demonstrate that the GPU-parallel algorithm that we developed improves the efficiency, accuracy and reliability of the outliers rejection in TESS light curves.

We perform a comparative study of different supervised machine learning time-series methods for short-term and long-term temperature forecasts on a real world dataset for the daily maximum temperature over North America given by DayMET. DayMET showcases a stochastic and high-dimensional spatio-temporal structure and is available at exceptionally fine resolution (a 1 km grid). We apply projection-based reduced order modeling to compress this high dimensional data, while preserving its spatio-temporal structure. We use variants of time-series specific neural network models on this reduced representation to perform multi-step weather predictions. We also use a Gaussian-process based error correction model to improve the forecasts from the neural network models. From our study, we learn that the recurrent neural network based techniques can accurately perform both short-term as well as long-term forecasts, with minimal computational cost as compared to the convolution based techniques. We see that the simple kernel based Gaussian-processes can also predict the neural network model errors, which can then be used to improve the long term forecasts.

We present SimNet, an AI-driven multi-physics simulation framework, to accelerate simulations across a wide range of disciplines in science and engineering. Compared to traditional numerical solvers, SimNet addresses a wide range of use cases - coupled forward simulations without any training data, inverse and data assimilation problems. SimNet offers fast turnaround time by enabling parameterized system representation that solves for multiple configurations simultaneously, as opposed to the traditional solvers that solve for one configuration at a time. SimNet is integrated with parameterized constructive solid geometry as well as STL modules to generate point clouds. Furthermore, it is customizable with APIs that enable user extensions to geometry, physics and network architecture. It has advanced network architectures that are optimized for high-performance GPU computing, and offers scalable performance for multi-GPU and multi-Node implementation with accelerated linear algebra as well as FP32, FP64 and TF32 computations. In this paper we review the neural network solver methodology, the SimNet architecture, and the various features that are needed for effective solution of the PDEs. We present real-world use cases that range from challenging forward multi-physics simulations with turbulence and complex 3D geometries, to industrial design optimization and inverse problems that are not addressed efficiently by the traditional solvers. Extensive comparisons of SimNet results with open source and commercial solvers show good correlation. The SimNet source code is available at https://developer.nvidia.com/simnet .

The present paper addresses the numerical fluid-structure interaction (FSI) analysis of a thermowell immersed in a water flow. The study was carried out implementing a modal superposition approach into a computational fluid dynamics (CFD) solver. The core of the procedure consists in embedding the structural natural modes, computed by a finite element analysis (FEA), by means of a mesh morphing tool based on radial basis functions (RBF). In order to minimize the distortion during the morphing action and to obtain a high quality of the mesh, a set of corrective solutions, that allowed the achievement of a sliding morphing on the duct surface, was introduced. The obtained numerical results were compared with experimental data, providing a satisfying agreement and demonstrating that the modal approach, with an adequate mesh morphing setup, is able to tackle unsteady FSI problems with the accuracy needed for industrial applications.

The increasing importance of optimization in manufacturing processes led to the improvement of well established optimization techniques and to the development of new and innovative approaches. Among these, an approach that exploits surface stresses distribution to obtain an optimized configuration is the Biological Growth Method (BGM). Coupling this method with surface sculpting based on Radial Basis Functions (RBF) mesh morphing had proven to be efficient and effective in optimizing specific mechanical components. In this work, the automatic, meshless and constrained parameter-less optimization approach is applied to a classical mechanical component and then compared with a parameter-based shape optimisation result.

Efficient execution of large-scale and extremely demanding computational scenarios is a challenge for both the infrastructure providers and end-users, usually scientists, that need to develop highly scalable computational codes. Nevertheless, at this time, on the eve of exa-scale supercomputers, the particular role has to be given also to the intermediate software that can help in the preparation of applications so they can be efficiently executed on the emerging HPC systems. The efficiency and scalability of such software can be seen as priorities, however, these are not the only elements that should be addressed. Equally important is to offer software that is elastic, portable between platforms of different sizes, and easy to use. Trying to fulfill all the above needs we present QCG-PilotJob, a tool designed to enable flexible execution of numerous potentially dynamic and interdependent computing tasks in a single allocation on a computing cluster. QCG-PilotJob is built on many years of collaboration with computational scientists representing various domains and it responses to the practical requirements of real scientific use-cases. In this paper, we focus on the recent integration of QCG-PilotJob with the EasyVVUQ library and its successful use for Uncertainty Quantification workflows of several complex multiscale applications being developed within the VECMA project. However, we believe that with a well-thought-out design that allows for fully user-space execution and straightforward installation, QCG-PilotJob may be easily exploited in many other application scenarios, even by inexperienced users.

Multiscale simulations present a new approach to increase the level of accuracy in terms of forced displacement forecasting, which can help humanitarian aid organizations to better plan resource allocations for refugee camps. People’s decisions to move may depend on perceived levels of safety, accessibility or weather conditions; simulating this combination realistically requires a coupled approach. In this paper, we implement a multiscale simulation for the South Sudan conflict in 2016–2017 by defining a macroscale model covering most of South Sudan and a microscale model covering the region around the White Nile, which is in turn coupled to weather data from the Copernicus project. We couple these models cyclically in two different ways: using file I/O and using the MUSCLE3 coupling environment. For the microscale model, we incorporated weather factors including precipitation and river discharge datasets. To investigate the effects of the multiscale simulation and its coupling with weather data on refugees’ decisions to move and their speed, we compare the results with single-scale approaches in terms of the total validation error, total execution time and coupling overhead.

Air pollution exposure is a major environmental risk to health and has been estimated to be responsible for 7 million premature deaths worldwide every year. This is of special concern in cities, where there are high levels of pollution and high population densities. Not only is there an urgent need for cities to monitor, analyze, predict and inform residents about the air quality, but also to develop tools to help evaluate mitigation strategies to prevent contamination. In this respect, the Weather Research and Forecasting model coupled with chemistry (WRF-Chem) is useful in providing simulations of meteorological conditions but also of the concentrations of polluting species. When combined with the multi-layer urban scheme Building Effect Parameterization (BEP) coupled with the Building Energy Model (BEM), we are furthermore able to include urban morphology and urban canopy effects into the atmosphere that affect the chemistry and transport of the gases. However, using WRF-Chem+BEP/BEM is computationally very expensive especially at very high urban resolutions below 5 km. It is thus indispensable to properly analyze the performance of these models in terms of execution time and quality to be useful for both operational and reanalysis purposes. This work represents the first step towards this overall objective which is to determine the performance (in terms of computational time and quality of results) and the scalability of WRF-BEP/BEM. To do so, we use the case study of Metropolitan Area of Barcelona and analyze a 24-h period (March 2015) under two with different Urban schemes (Bulk and BEP/BEM). We analyze the execution time by running the two experiments in its serial configuration and in their parallel configurations using 2, 4, 8, 16, 32 and 64 cores. And the quality of the results by comparing to observed data from four meteorological stations in Barcelona.

Neurodegenerative diseases are frequently associated with aggregation and propagation of toxic proteins. In particular, it is well known that along with amyloid-beta, the tau protein is also driving Alzheimer’s disease. Multiscale reaction-diffusion models can assist in our better understanding of the evolution of the disease. Based on a coarse-graining procedure of the continuous model and taking advantage of the brain data connectome, a computationally challenging network mathematical model has been described where the edges of the network are the axonal bundles in white-matter tracts. Further, we have modified the heterodimer model in such a way that it can now capture some of the critical characteristics of this evolution such as the conversion time from healthy to toxic proteins. Finally, we have analyzed the modified model theoretically and validated the theoretical findings with numerical simulations.

In this paper, we present an efficient numerical method for solving the electrical activity of the heart. We propose a second order alternating direction implicit finite difference method (ADI) for both space and time. The derivation of the proposed ADI scheme is based on the semi-implicit backward differentiation formula (SBDF). Numerical simulation showing the computational advantages of the proposed algorithm in terms of the computational time and memory consumption are presented.

Upscaling of the mechanical properties of polycrystalline aggregates might require complex and time-consuming procedures, if adopted to help in the design and reliability analysis of micro-devices. In inertial micro electro-mechanical systems (MEMS), the movable parts are often made of polycrystalline silicon films and, due to the current trend towards further miniaturization, their mechanical properties must be characterized not only in terms of average values but also in terms of their scattering. In this work, we propose two convolutional network models based on the ResNet and DenseNet architectures, to learn the features of the microstructural morphology and allow automatic upscaling of the statistical properties of the said film properties. Results are shown for film samples featuring different values of a length scale ratio, so as to assess accuracy and computational efficiency of the proposed approach.

In this paper, we present the framework for the multiscale thermoelastic analysis of composites. Asphalt concrete (AC) was selected to demonstrate the applicability of the proposed approach. It is due to the observed high dependence of this material performance on the thermal effects. The insight into the microscale behavior is upscaled to the macroresolution by the multiscale finite element method (MsFEM) that has not been used so far for coupled problems. In the paper, we present a brief description of this approach together with its new application to coupled thermoelastic numerical modeling. The upscaled results are compared with the reference ones and the error analysis is presented. A very good agreement between these two solutions was obtained. Simultaneously, a large reduction of the degrees of freedom can be observed for the MsFEM solution. The number of degrees of freedom was reduced by 3 orders of magnitude introducing an additional approximation error of only about 6%. We also present the convergence of the method with the increasing approximation order at the macroresolution. Finally, we demonstrate the impact of the thermal effects on the displacements in the analyzed asphalt concrete sample.

When a molecular dynamics (MD) simulation and a computational fluid dynamics (CFD) solver are coupled together to create a multiscale, molecular-continuum flow simulation, thermal noise fluctuations from the particle system can be a critical issue, so that noise filtering methods are required. Noise filters are one option to significantly reduce these fluctuations.We present a modified variant of the Non-Local Means (NLM) algorithm for MD data. Originally developed for image processing, we extend NLM to a space-time formulation and discuss its implementation details.The space-time NLM algorithm is incorporated into the Macro-Micro-Coupling tool (MaMiCo), a C++ molecular-continuum coupling framework, together with a novel flexible filtering subsystem. The latter can be used to configure and efficiently execute arbitrary data-flow chains of simulation data analytics modules or noise filters at runtime on an HPC system, even including python functions. We employ a coupling to a GPU-based Lattice Boltzmann solver running a vortex street scenario to show the benefits of our approach. Our results demonstrate that NLM has an excellent signal-to-noise ratio gain and is a superior method for extraction of macroscopic flow information from noisy fluctuating particle ensemble data.