AI-Newton: Concept-Driven Physical Law Discovery System Without Prior Knowledge (arxiv.org)

AI-Newton is a concept-driven discovery system that autonomously derives physical laws from raw, noisy experimental data without supervision or any prior physical knowledge. The authors present a knowledge base and a concept-centered representation plus an autonomous discovery workflow that invents and assembles intermediate concepts, hypotheses and symbolic relations. As a proof of concept, AI-Newton was tested on a large suite of Newtonian mechanics problems and successfully re-discovered canonical laws — including Newton’s second law, energy conservation, and the law of gravitation — by defining concepts and relations internally rather than being told variables or equations beforehand. This work is significant because it moves beyond pattern-fitting toward machine-led scientific reasoning: the system can generate interpretable, symbolic laws from data alone, tolerates noise, and structures discovery around reusable concepts. Technically, the contribution lies in coupling a concept-focused knowledge representation with an autonomous search over symbolic hypotheses (rather than relying on human-specified priors), enabling scalable, domain-agnostic law discovery. While demonstrated in classical mechanics, AI-Newton points to broader implications for automated hypothesis generation, interpretable AI in science, and accelerated discovery pipelines—while still requiring further validation on more complex or quantum/thermodynamic domains.