Benchmarking of Deep Learning models on 2D Laminar Flow behind Cylinder

TL;DR

This paper evaluates deep learning models, especially Transformers, for simulating 2D laminar flow behind a cylinder, demonstrating that Transformers outperform other architectures in this fluid dynamics task.

Contribution

It introduces the application of Transformer architectures to Direct Numerical Simulation in CFD, showing their superior performance over traditional deep learning models.

Findings

Transformers outperform other models on the DNS dataset.

Deep learning models can effectively simulate fluid flow in CFD.

Transformers are promising for future CFD research.

Abstract

The rapidly advancing field of Fluid Mechanics has recently employed Deep Learning to solve various problems within that field. In that same spirit we try to perform Direct Numerical Simulation(DNS) which is one of the tasks in Computational Fluid Dynamics, using three fundamental architectures in the field of Deep Learning that were each used to solve various high dimensional problems. We train these three models in an autoencoder manner, for this the dataset is treated like sequential frames given to the model as input. We observe that recently introduced architecture called Transformer significantly outperforms its counterparts on the selected dataset.Furthermore, we conclude that using Transformers for doing DNS in the field of CFD is an interesting research area worth exploring.