AirfRANS / neural CFD surrogate

This fake writeup pretends the goal is simple: make a neural surrogate that behaves enough like a CFD solver to be useful, but not so much like a CFD solver that it takes all afternoon to answer.

The model gets geometry, flow conditions, and a stern lecture from validation metrics. It returns fields, uncertainty hints, and the occasional reminder that conservation laws are not optional.

The interesting part is the boundary between learned approximation and solver-grounded truth: where the network is fast, where it is wrong, and how to know before a wing falls off in a slide deck.