Scholarlink

Abstract <p>The Viscous Vortex Domains (VVD) method is a powerful fully Lagrangian approach for simulating two-dimensional viscous incompressible flows. However, its computational efficiency is severely hampered by the high cost of calculating the diffusive velocity field, which is crucial for modeling viscous effects. Specifically, the evaluation of contour integrals for terms accounting for the influence of solid boundaries becomes the most time-consuming part of the algorithm, consuming up to 70% of the total simulation time, since <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(N\)</EquationSource> <!--TechPhys2670029Sukhova-m1--> </InlineEquation>-body-type fast methods that allow for a “dramatic” speedup of the rest of the algorithm turn out to be inefficient for this operation. This paper presents a novel approach to accelerate this calculation by replacing the direct numerical integration with a fast approximation using a pre-trained multilayer perceptron (MLP). The neural network approximates two dimensionless functions, which are universal for all panels, based on three normalized geometric parameters. The model was trained on a high-accuracy dataset generated in Wolfram Mathematica. Implemented within the VM2D code using NVIDIA’s cuDNN library for GPU acceleration, this approach demonstrates a significant performance improvement. The computation time for the diffusive velocity terms was reduced by approximately 85%, leading to a 1.75x overall speedup of the simulation time step. Furthermore, the new method provides higher accuracy, resulting in a cleaner, more periodic force signal and improved simulation quality, as demonstrated for flow around an impulsively started cylinder at <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({\text{Re}} = 1000\)</EquationSource> <!--TechPhys2670029Sukhova-m2--> </InlineEquation>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Viscous Vortex Domains Method Implementation with Neural Network Approach for Performance and Accuracy Improvement

  • 期刊论文

  • 发表日期：

    2026年5月23日

查看全文
  • E. Sukhova,
  • I. Marchevsky,
  • I. Kazey

摘要

Abstract

The Viscous Vortex Domains (VVD) method is a powerful fully Lagrangian approach for simulating two-dimensional viscous incompressible flows. However, its computational efficiency is severely hampered by the high cost of calculating the diffusive velocity field, which is crucial for modeling viscous effects. Specifically, the evaluation of contour integrals for terms accounting for the influence of solid boundaries becomes the most time-consuming part of the algorithm, consuming up to 70% of the total simulation time, since \(N\) -body-type fast methods that allow for a “dramatic” speedup of the rest of the algorithm turn out to be inefficient for this operation. This paper presents a novel approach to accelerate this calculation by replacing the direct numerical integration with a fast approximation using a pre-trained multilayer perceptron (MLP). The neural network approximates two dimensionless functions, which are universal for all panels, based on three normalized geometric parameters. The model was trained on a high-accuracy dataset generated in Wolfram Mathematica. Implemented within the VM2D code using NVIDIA’s cuDNN library for GPU acceleration, this approach demonstrates a significant performance improvement. The computation time for the diffusive velocity terms was reduced by approximately 85%, leading to a 1.75x overall speedup of the simulation time step. Furthermore, the new method provides higher accuracy, resulting in a cleaner, more periodic force signal and improved simulation quality, as demonstrated for flow around an impulsively started cylinder at \({\text{Re}} = 1000\) .