6.13 AI Feynman as a controlled testbed

Symbolic regression is an unusually clean meeting point between machine learning and scientific discovery. Given samples from an unknown function, the system must recover a symbolic expression rather than merely interpolate the samples. Udrescu and Tegmark observe that scientifically interesting functions often possess simplifying structure—symmetry, separability, compositionality, dimensional constraints, and simple transformations—even though unrestricted symbolic regression is combinatorially hard [ Udrescu and Tegmark , 2020 ] .

AI Feynman combines neural fitting with physics-inspired tests that detect and recursively exploit this structure. On the authors’ benchmark of one hundred equations drawn from the Feynman Lectures on Physics, it recovered all one hundred, compared with seventy-one for the previously available systems in their comparison. The important fact for this book is not only the score. The algorithm succeeds by diagnosing structural properties and using them to factor a hard search into simpler subproblems.

The name also recalls Feynman’s atomic-hypothesis thought experiment [ Feynman et al. , 1963 , Section 1–2 ] . Its ideal is not the isolated recovery of one equation, but a compact theory whose primitives and laws make many phenomena intelligible. Equation recovery is thus a natural controlled baseline, while recovery of the generative theory remains the more ambitious target.

Categorically, a recovered formula can be read as a term in a many-sorted algebraic theory. Units constrain sorts; composition records the expression tree; symmetry identifies invariant actions; and separability exposes a factorization diagram. AI Feynman therefore supplies a concrete benchmark for the claim that compositional structure can guide discovery.

It is not yet a benchmark for full theory construction. The variables are given, the observations are generated by a known closed-form expression, the language of candidate formulas is largely fixed, and success is defined by recovering an existing law. This is primarily exploratory creativity within a carefully designed expression space. A synthetic-creativity benchmark should preserve that controlled core while progressively relaxing those supports.