JAX-based differentiable fluid dynamics on GPU and end-to-end optimization

TL;DR

This paper introduces JAX-Fluids, a GPU-accelerated differentiable fluid dynamics solver that leverages automatic differentiation for efficient high-dimensional optimization, validated on turbulent and porous media flows.

Contribution

The development of JAX-Fluids, a novel differentiable fluid dynamics solver based on JAX, enabling end-to-end optimization on GPU with validated accuracy and performance.

Findings

JAX-Fluids achieves performance comparable to high-order codes like FLEXI.

Validated with DNS of turbulent channel flow showing excellent agreement.

Successfully tested a new boundary condition for porous media flows.

Abstract

This project aims to advance differentiable fluid dynamics for hypersonic coupled flow over porous media, demonstrating the potential of automatic differentiation (AD)-based optimization for end-to-end solutions. Leveraging AD efficiently handles high-dimensional optimization problems, offering a flexible alternative to traditional methods. We utilized JAX-Fluids, a newly developed solver based on the JAX framework, which combines autograd and TensorFlow's XLA. Compiled on a HAWK-AI node with NVIDIA A100 GPU, JAX-Fluids showed computational performance comparable to other high-order codes like FLEXI. Validation with a compressible turbulent channel flow DNS case showed excellent agreement, and a new boundary condition for modeling porous media was successfully tested on a laminar boundary layer case. Future steps in our research are anticipated.