Scientific Ml

Hybrid Physics-ML Models

Hybrid models combine a trusted physical model with a learned component that corrects its errors or supplies its missing pieces.

The best of both

Pure physics models are reliable but incomplete or coarse; pure data models are flexible but ignorant of known laws. A hybrid physics-ML model keeps the physics for what it does well and adds a learned component for what it does poorly. The result generalizes better than a black box and is more accurate than physics alone, using less data than either extreme.

Common patterns

Kronos motion — pid vs model

Residual correction: physics gives a baseline, a network learns the leftover error
Parameter learning: a network infers uncertain coefficients the physics then uses
Closure modeling: a network supplies an unresolved subgrid term to a coarse solver
Physics-guided loss: a data model is penalized for violating known equations

Residual learning

The most common pattern predicts the discrepancy between a physical model and reality. Because the physics captures the bulk of the behavior, the residual is small and structured, so a modest network learns it from limited data. This keeps the model anchored to physics: where the network has no data, the prediction defaults to the physical baseline rather than to nonsense.

Keeping physics dominant

A well-designed hybrid ensures the learned part cannot override the physics arbitrarily. Bounding the correction, regularizing it toward zero, and enforcing conservation on the combined output all keep the physics in charge. This matters most under extrapolation, where the data-driven part is least trustworthy and the physical prior is the safeguard.

Application in modeling

Hybrid models are studied for turbulence closures, chemical kinetics, and plasma transport, where first-principles terms are known but expensive subscale effects must be modeled. For design machines such as the Hyperion breeder, a hybrid surrogate can accelerate repeated transport evaluations while remaining tethered to the underlying equations, and its outputs are validated against full solvers before use.