Diff-FlowFSI: A GPU-Optimized Differentiable CFD Platform for Scalable Forward and Inverse Modeling of Complex Flows

Turbulent flow and fluid–structure interaction (FSI) problems are central to many scientific and engineering domains, yet remain computationally demanding due to their nonlinear, multiscale, and coupled nature. While traditional CFD solvers offer high-fidelity simulations, they are often ill-suited for modern scientific workflows that require scalability, differentiability, and integration with learning-based components for inverse modeling, optimization, and data assimilation.

We present Diff-FlowFSI, a GPU-accelerated, fully differentiable CFD platform for scalable forward and inverse modeling of complex flow and FSI systems. Built in JAX, we built a fully vectorized finite volume solver with an immersed boundary method to support arbitrarily complex geometries and strong fluid–structure coupling. Its differentiable architecture enables end-to-end gradient computation via automatic differentiation, facilitating hybrid neural–physics modeling, gradient-based optimization, and deep learning integration. We demonstrate the platform’s accuracy, efficiency, and scalability through a suite of canonical turbulence and FSI benchmarks, positioning Diff-FlowFSI as a powerful tool for scientific machine learning and differentiable simulation at scale.