Hybrid solver with local correction using Lagrangian latent memory

Gwendal Jouan, Matthias Schulz, Daniel Berger, Stefan Gavranovic, Dirk Hartmann

Keywords: differentiable physics, hybrid methods, solver-in-the-loop

TL;DR: Hybrid CFD solver with local learned correction using transported latent variables

Abstract: Augmentation of differentiable CFD solvers with neural networks has shown promising results. However, most approaches rely on convolutional neural networks (CNN) and Cartesian solvers with efficient access to all cell data. This choice poses challenges for industrial solvers that operate on unstructured meshes and with efficient access to neighboring cells only. In this work, we address this limitation using a novel architecture, named Transported Memory Networks. The architecture draws inspiration from both traditional turbulence models and neural networks and is compatible with generic discretizations. We demonstrate that it is point-wise and statistically comparable to, or improves upon, previous methods in terms of accuracy and efficiency.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 30