Machine learning excels at finding patterns and making predictions but generally cannot convert either into general scientific concepts. A model can predict planetary orbits without arriving at a law of gravitation, leaving humans to interpret the equations and extract the principle. A team at the School of Physics at Peking University built AI-Newton, a system designed to discover fundamental physics laws from experimental data without human guidance.
The architecture reflects that goal. Rather than fitting one model to one dataset, AI-Newton builds a knowledge base of concepts and laws incrementally, an approach the team describes as mimicking the cognitive process of human scientists. It uses symbolic regression, which searches for the mathematical expression that best describes a phenomenon, so its outputs are equations that can be inspected and reused rather than opaque parameter sets.
The reuse is what distinguishes it. In one demonstration the system was given data on a ball's position over time and derived an equation for its velocity. It stored that result and drew on it in a later task to calculate the ball's mass using Newton's second law, building a derived quantity on top of a previously discovered one. The system was tested across 46 physics experiments, and Nature published a news article about the work.
The group is now testing whether the approach can uncover quantum theories, a considerably harder target where the relevant concepts are less intuitively available from raw measurement. The broader interest for philosophers of science as much as physicists is whether cumulative, concept-forming discovery can be automated, or whether the formation of new concepts remains distinctively human.