“AI Feynman: a Physics-Inspired Method for Symbolic Regression,” by Silviu-Marian Udrescu and Max Tegmark at MIT, was submitted to arXiv in May 2019 and published in Science Advances in 2020. It tackled symbolic regression: given a table of input-output data, find the actual mathematical formula that generated it, expressed in symbols like sin, exp and division, rather than a black-box numerical fit.
Symbolic regression is hard because the space of possible formulas is enormous. The authors built a recursive algorithm that combines a neural network with techniques borrowed from physics, exploiting properties such as symmetry, separability and dimensional analysis to break a complicated equation into simpler pieces the method can solve. Tested on 100 equations drawn from the Feynman Lectures on Physics, AI Feynman recovered all 100, where the best previous software cracked only 71. On a harder set of equations, it raised the success rate from 15 percent to 90 percent.
The work matters because much of science is ultimately about finding compact laws that explain data, and a tool that can propose interpretable formulas, not just predictions, points toward AI that helps discover new physics rather than merely curve-fitting. It became one of the most cited demonstrations of machine-assisted scientific discovery.