Transolver: A Transformer Solver for Large-Scale Industrial Geometries

Standard attention over millions of mesh points is prohibitively expensive. Transolver instead learns a slice-based projection that groups mesh points into a modest number of learnable physical states, and performs attention in that compressed space.

Reported gains: State-of-the-art results across standard PDE benchmarks and industrial car/aerodynamics cases, with substantially lower memory than plain Transformers.