Overview
A neural flow-map for incompressible turbulent flows, trained on Nek5000 simulations. The architecture learns a one-step push-forward operator and is composed at inference time to produce long-horizon predictions.
Method
The training objective combines a primary L2 step loss with a Sobolev penalty on velocity gradients and a noise-injection term that improves stability under autoregressive rollout. We adopt a curriculum that grows the rollout horizon as accuracy improves.
Status
In progress as of May 2026. Master thesis at the MIT van Rees Lab.