"Recursive Flow Matching"

Recursive Flow Matching makes physics forecasting much faster without losing accuracy.

So most diffusion models need many steps to simulate systems like fluids, climate, or wave dynamics. While vanilla flow matching is faster, it still often breaks down when forced into 1 or 2 steps.

So this paper fixes this by training the model on multiple rescaled flow paths that meet at the same state, then making those paths agree.

Which makes the learned trajectory smoother and more stable, so the model can generate accurate rollouts in just 1 or 2 steps.

It gets up to 20x faster inference than diffusion-based emulators and over 15% lower MSE than vanilla flow matching.