AI learns to predict expensive fluid flow physics in seconds instead of hours

Researchers built a vision transformer model that can predict complex fluid flows (like gas injection in engines) by learning from CFD simulation data instead of running simulations from scratch. Instead of waiting hours for expensive computational physics, the model produces predictions in seconds and can reconstruct missing data from partial observations.

Computational fluid dynamics simulations are a bottleneck in engineering: they're expensive, slow, and required for every new design variation. If this approach scales from the lab to industry workflows, design iteration becomes faster and cheaper. The question is whether companies actually adopt it — faster physics predictions only matter if they change how engineers work, not just how fast a model can run in isolation.