Surrogate Solvers

Fast learned surrogates for expensive physics simulations

Overview

Learned surrogate models standing in for expensive numerical solvers (field simulations, plasma dynamics) — trading a controlled amount of accuracy for orders-of-magnitude speedup, with error bounds tracked explicitly rather than ignored.

Architecture

• Learned surrogate models standing in for expensive numerical solvers (field simulations, plasma dynamics)
• Error bounds tracked explicitly rather than ignored — the surrogate reports its own uncertainty, not just a point estimate
• Natural first target: tardigrade-cad's Biot-Savart solver, since it already exists and has a clear ground truth to train against

Milestones

1. 01Pick one solver to surrogate first — tardigrade-cad's Biot-Savart solver is the leading candidate
2. 02Establish an accuracy/speed tradeoff curve against the original solver
3. 03Integrate error bounds into a downstream consumer, e.g. the tardigrade-cad optimization loop

Open questions

• Whether Biot-Savart or a BOUT++ plasma quantity should be the first surrogate target
• How error bounds propagate into downstream design decisions — if the surrogate is wrong, does the optimizer know how wrong?