Equation Discovery
Equation discovery infers the governing laws of a system directly from data, aiming for compact, physically meaningful models.
The goal
Equation discovery seeks the mathematical laws that generated observed data, not just a predictor. Where a black-box model answers only what happens next, a discovered equation explains why, exposing the structure of the dynamics in a form a scientist can inspect, test, and connect to theory.
Families of methods
- Sparse regression over a candidate library, as in SINDy, for differential equations
- Symbolic regression via genetic programming for free-form algebraic laws
- Neural approaches that learn a differentiable model then distill it to symbols
- Grammar- or physics-guided search that restricts candidates to dimensionally valid forms
The role of priors
Data alone rarely single out one law; many expressions fit a finite sample. Physical priors break the tie. Dimensional consistency rejects expressions with mismatched units. Known symmetries and conservation laws prune whole regions of the search space. Requiring simplicity favors laws that generalize. These priors are what make discovery reliable rather than a curve-fitting exercise.
Validation
A discovered equation must earn trust. The strongest test is extrapolation: a genuine law predicts regimes absent from the training data, while an overfit expression fails there. Reproducibility across independent datasets and stability across noise realizations are further checks. A candidate that survives these is a hypothesis worth deeper study, not a settled truth.
Where it helps
Equation discovery is valuable when the mechanism is unknown or a fast reduced model is needed for a well-understood but expensive system. In plasma and fusion research it is explored for finding compact closures and reduced transport relations from simulation and experiment, always cross-checked against first-principles models rather than replacing them.