AI Feynman: A Physics-Inspired Method for Symbolic Regression (arxiv.org)

Researchers introduced AI Feynman, a physics-inspired symbolic regression method that marries neural-network fitting with a toolbox of physics heuristics (dimensional analysis, symmetry detection, variable separability, functional composition and recursive decomposition) to recover closed-form expressions from data. Instead of brute-force search over expression space, the algorithm fits a smooth neural approximation, uses that to test for simplifications (e.g., separability, invariances, useful variable transformations), and then recursively reduces the problem to simpler subproblems that can be solved by targeted brute force or symbolic search. The authors released code, data and demos. The approach is significant because it shows how domain priors dramatically shrink an otherwise intractable (likely NP-hard) search and produce interpretable laws from noisy samples. On benchmark tests drawn from the Feynman Lectures, AI Feynman recovered all 100 target equations (vs 71 by prior public tools) and boosted success on a harder test set from 15% to 90%. For the AI/ML community this underscores a practical path to combining flexible function approximators with structured, theory-driven search to discover symbolic models—useful for scientific discovery, model compression, and interpretable ML pipelines.