Compute-optimal data scaling for neural surrogates via multi-fidelity training
Setinek et al.
What it does
Treats simulation fidelity as a spectrum and finds the mix of cheap and accurate training samples that minimises error for a fixed data-generation budget, across six PDEs, two error sources and four architectures.
Why it matters
The right fidelity mix matches high-fidelity-only accuracy for about 3× less simulation compute on average (up to 6.3×), and lowering solver tolerance is the easiest place to start.