A CNN-based particle tracking method for large-scale fluid simulations with Lagrangian-Eulerian approaches

TL;DR

This paper introduces a CNN-based particle tracking method that significantly reduces computational costs and enhances scalability in large-scale fluid simulations using Lagrangian-Eulerian approaches.

Contribution

It proposes the CNN-SNS method that shortens tracking paths and reduces computational overhead, improving efficiency over traditional neighbor search methods.

Findings

Achieves 95.8% reduction in tracking path length

Achieves 97.0% reduction in computational time

Demonstrates superior scalability in high-velocity, large-scale flows

Abstract

A novel particle tracking method based on a convolutional neural network (CNN) is proposed to improve the efficiency of Lagrangian-Eulerian (L-E) approaches. Relying on the successive neighbor search (SNS) method for particle tracking, the L-E approaches face increasing computational and parallel overhead as simulations grow in scale. This issue arises primarily because the SNS method requires lengthy tracking paths, which incur intensive inter-processor communications. The proposed method, termed the CNN-SNS method, addresses this issue by approximating the spatial mapping between reference frames through the CNN. Initiating the SNS method from CNN predictions shortens the tracking paths without compromising accuracy and consequently achieves superior parallel scalability. Numerical tests demonstrate that the CNN-SNS method exhibits increasing computational advantages over the SNS method in large-scale, high-velocity flow fields. As the resolution and parallelization scale up, the CNN-SNS method achieves reductions of 95.8% in tracking path length and 97.0% in computational time.