Scientific Ml

Multi-Fidelity Modeling

Multi-fidelity methods combine many cheap approximate runs with a few expensive accurate ones to build an accurate model efficiently.

Not all runs cost the same

A physical system can often be simulated at several levels of fidelity: a coarse mesh or simplified physics is cheap but approximate, while a fine mesh or full physics is accurate but expensive. Multi-fidelity modeling exploits this hierarchy, using many cheap low-fidelity runs to learn the general shape of the response and a few costly high-fidelity runs to correct it.

How the levels combine

Kronos motion — pid vs model

A common construction models the high-fidelity output as a transformed low-fidelity output plus a learned correction. The low-fidelity model, trained on abundant data, captures the trends; the correction, trained on scarce high-fidelity data, fixes the systematic bias. Because the correction is small and smooth, it can be learned from few points, so the accurate model costs far less than training on high-fidelity data alone.

Methods

Co-kriging: a Gaussian process framework linking correlated fidelity levels
Multi-fidelity neural networks: separate subnetworks for trend and correction
Control variates: use cheap runs to reduce the variance of expensive estimates
Hierarchies of more than two levels for finely graded cost and accuracy

When it pays off

Multi-fidelity modeling helps most when the cheap and expensive models are well correlated, so the low-fidelity output genuinely predicts the high-fidelity trend. If the correlation is weak, the cheap runs add little and the method degrades toward using high-fidelity data alone. Checking the correlation between fidelities is the first step before committing to the approach.

Use in engineering

Design studies routinely have a ladder of models, from fast reduced physics to slow high-resolution simulation. Multi-fidelity methods let a study lean on the cheap models while spending its scarce expensive evaluations where they matter, delivering accuracy that neither level could provide economically on its own.