Autonomous Discovery of Particle Physics Theories from Data

by Loukas Gouskos (Brown University (US))

Thursday 3 Sept 2026, 13:30 → 14:30 Europe/Zurich

4/2-037 - TH meeting room (CERN)

Dark matter and the matter-antimatter asymmetry, among other observations, indicate that the Standard Model is incomplete, yet decades of experiments have found no direct evidence of the new physics responsible. The space of possible theories is vast and we can explore only a few models at a time, making the search for the one realized in Nature a formidable challenge. We present Albert, a neuro-symbolic framework that constructs quantum field theories directly from data. Symmetries, particle content, and interactions are generated as sequences under a formal grammar, so that every candidate is a well-defined and internally consistent Lagrangian by construction. A reinforcement learning loop enforces theoretical constraints, computes observables with radiative corrections, and scores candidates by their χ^2 agreement with measured data. As a proof of concept, a 25M-parameter model trained only on legacy LEP data, which contains no direct evidence of the top quark, rediscovers it and recovers its mass from precision electroweak observables alone. We discuss the approach and its prospects for a systematic, data-driven exploration of BSM theory space.