Supercomputing Frontiers and Innovations

Simulation of Seismic Processes with High-Order Grid-Characteristic Methods

2024-11-05T13:47:34+05:00

This paper considers simulation of the seismic wave propagation in geological media with different rheological properties. The present work aims to construct a numerical scheme to model porous fluid-saturated medium, for the description of which the Dorovsky model was selected. We employed the grid-characteristic method, which includes choosing the appropriate operator splitting method for a 3D problem, deriving the transformation to the Riemann invariants analytically, and explicitly setting boundary and contact conditions. We simulated two scenarios. Firstly, we compared the wavefields generated by a point-source in the acoustic, linear elastic, and porous fluid-saturated approximations, noting the similarities in the longitudinal wave and differences in other wave types. Secondly, we simulated a part of the marine seismic survey process, including a source in the water layer, governed by the acoustic equations, a water-saturated layer described by the Dorovsky equations, and an explicit contact between these layers. To utilize the modern HPC multi-core and multi-processor systems, the hybrid MPI-OpenMP parallel algorithms were used.

Numerical Modeling of Complex Geometry Thin Composite Structures under Vibrational Testing

2024-11-09T02:14:39+05:00

This paper considers the inverse problem of material elastic properties identification from vibrational testing data. The present work aims to describe the approach that uses different kinds of optimizations to allow fast inverse problem solution using single modern multicore CPU or GPU. This includes choosing the model that allows to minimize the computational cost still reproducing the experimental results with good quality. The model for mid-surface symmetric isotropic and composite plates that are moving in vibrational stand is provided. The inverse problem is formulated in terms of the loss function minimization and the solution is computed with stochastic global optimization algorithm and second-order local optimization algorithm, which uses automatic differentiation of the forward problem solver to compute the derivatives. The paper describes parallelization for CPU and GPU and also the approach to reduce RAM usage to fit into single server RAM or single GPU VRAM. The numerical experiments presented in the paper demonstrate the solutions for complex rheologies and geometries: laminated composite plates, isotropic materials with frequency dependent elastic properties, perforated samples.

Numerical Modeling of Marine Seismology in the Arctic Region During Deposit Dissolution due to Oil Migration

2024-11-09T02:21:19+05:00

As the Arctic region requires costly and difficult to access surveys of the ground, numerical modeling proves to be an effective way to study occurring processes in the area. Moreover, the simulations of the seismic exploration can help identify the main signs of oil migration which is crucial for the risk assessment of the deposit development. Therefore, the main goal of this work is to conduct the forward modeling of the seismology in the offshore areas of the region in order to determine the indicators of such processes. The present study, in particular, is aimed at recreating of basic features of the region such as a layered ground with the gradual change in material parameters and inclusion of a permafrost sheet. Furthermore, the boundaries between layers are considered to be curvilinear. This complex problem was effectively solved using the gridcharacteristic method which allows for the accelerating of the calculations using OpenMP. As a result of the computations, the reconstructed wave phenomena is analyzed based on the obtained wave patterns and synthesized seismograms. The change in the responses from the oil reservoir in the process of draining is identified which can further help interpret real measurements.

A Modification of Adaptive Greedy Algorithm for Solving Problems of Fractured Media Geophysics

2024-11-09T02:26:51+05:00

Nowadays, the issue of direct modeling of seismic exploration problems is becoming increasingly important due to the development of a new field of application of such algorithms as generation of a training samples for subsequent solution of the appropriate inverse problem using neural networks. This challenges scientists to develop corresponding parallel algorithms and improve their efficiency. The current manuscript is devoted to the algorithm for decomposing a large number of individual computational grids of various sizes for a large number of MPI processes using the example of a 3D direct problem of seismic exploration of geological media treating the complex topology of the Earth’s surface, the complex shape of interfaces between geological layers and a large number of explicitly treated geological fractures, that are not aligned with the coordinate axes. Three modifications of the grid-characteristic numerical method on Chimera and curvilinear computational grids are compared with each other. The dependence on different numbers of fractures is studied. A large number (several hundreds or thousands) of fractures in the geological media significantly increases the amount of transmitted data, which imposes requirements on the developed modification of the greedy algorithm.

On an Algorithm for Decomposing Multi-Block Structured Meshes for Calculating Dynamic Wave Processes in Complex Structures on Supercomputers with Distributed Memory

2024-11-05T13:58:11+05:00

The advancement of the oil and gas industry represents a key priority area for the Russian Federation. The Arctic region contains substantial hydrocarbon reserves, but the inherent difficulties in exploring these resources make them particularly challenging to access. The present paper is devoted to the numerical calculation of the dynamic impact propagation on an oil platform using parallel computing methods. To address this issue, a grid-characteristic method was employed. The substantial volume of computation necessitates the utilization of parallel computing techniques, such as Message Passing Interface (MPI). A grid model was constructed based on a real platform, and an algorithm for decomposing the computational domain was developed with the aim of reducing the message time between MPI processes and increasing speedup. A series of test calculations were performed to demonstrate the capabilities of the algorithms. Examples of calculations and the application of the developed method of decomposition are provided. The feasibility of decomposition and parallelization algorithms is currently being investigated. The conducted tests have demonstrated the potential for using the model for real calculations.

Parallelization Strategies for Ultrasonic Wave Propagation in Composite Materials Considering Microstructural Details

2024-11-05T13:58:14+05:00

This paper explores advanced parallelization strategies for simulating ultrasonic wave propagation in composite materials considering their complex microstructure. The grid-characteristic method and the use of Chimera grids in the simulations allow us to represent the composite material as an isotropic, linear-elastic medium and focus on improving the computational efficiency through efficient grid partitioning techniques.We used MPI (Message Passing Interface) technology on a high-performance computing cluster to test different methods for distributing computational grids across multiple processes. Our results highlight that partitioning grids according to material fiber layers improves the performance, especially when the number of processes matches the number of composite layers. This method not only provides better load balancing but also reduces communication overhead, making it the most efficient strategy tested. We present a comprehensive comparison of execution times and speedups for different partitioning approaches. Future work will aim to extend the study by increasing the number of layers and exploring how this approach scales with more complex and heterogeneous microstructures, potentially identifying further optimizations for parallel modeling.

Leveraging OpenMP Tasks for Efficient Parallel Modeling of the Elastic Eave Propagation in Multi-mesh Problems

2024-11-09T02:33:28+05:00

This paper presents a new algorithm for parallelizing the grid-characteristic method in sharedmemory systems. The OpenMP task parallelism mechanism is used for parallelization. A modification of the grid-characteristic method is considered that uses a set of overlapped grids to determine a complex heterogeneous structure of the computational domain. The complexity of parallelizing the algorithm is represented by the presence of many different-sized grids. The proposed algorithm is described and compared with basic parallelization algorithms. Basic algorithms mean separate parallelization within each computational grid using the loop parallelization mechanism. An analysis of the efficiency of the post-doubling and parallel algorithms is performed. The advantage of the proposed algorithm for a number of problems is demonstrated. The results of testing and calculating the propagation of wave disturbances in a fractured layer are presented. Each crack in the example is specified by a separate computational grid, which significantly increases the multi-scale problem and the number of computational grids. Work is underway to transfer the algorithm to the three-dimensional case.

HPC Optimization Algorithm for Assessing Haemodynamic Parameters in Synthetic Patient Cohorts

2024-11-18T12:50:08+05:00

A computational framework for the generation of a synthetic pulse wave database is developed. This framework demonstrates the feasibility of generating large-scale, high-fidelity virtual patient cohorts for biomedical research. Pulse waves are generated using a one-dimensional hemodynamic model of the systemic circulation coupled with a model of the left heart. Each virtual patient in the database is defined by a set of physiological parameters, including systolic and diastolic blood pressure, stroke volume, and heart rate. The parameters are optimized to match the desired outputs by solving an inverse problem using the Unscented Kalman Filter (UKF). The UKF is selected for its ability to accurately and efficiently estimate parameters in nonlinear systems. The generation of a single virtual patient requires between one and several hundred iterations of the UKF, depending on the complexity of the desired outputs. To meet the computational demands of generating a database with thousands of virtual patients, a computing cluster with 24 CPU nodes, each containing 52 cores, is employed. Two levels of parallelization are implemented, resulting in a speedup factor of 8.