Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers
Medrano-Navarro et al.
What it does
Pre-trains surrogates on unstructured 3D geometries without a stored dataset: from geometric shape descriptors for steady problems, and from synthetic PDE data generated online for transient ones.
Why it matters
Faster convergence and better fine-tuning accuracy when labelled simulations are few, without the disk and generation cost of a large pre-training corpus.