New AI framework can uncover space physics equations in raw data (phys.org)

Researchers from Tsinghua, Peking University and collaborators introduced PhyE2E, a neural-symbolic AI that automatically extracts compact, unit-consistent symbolic equations from raw space‑physics data (paper in Nature Machine Intelligence). Rather than just curve-fitting, PhyE2E learns a prior over plausible, dimensionally consistent physics expressions by fine-tuning on established equations and then translates observational data directly into symbolic formulas and their units using a transformer. Tested on synthetic LLM-generated datasets and real NASA records, the system recovered human-derived laws across five space‑physics scenarios and even produced an improved mathematical description of solar cycles, plus effective relations linking solar radiation, temperature and magnetic fields. Technically, PhyE2E breaks hard problems into simpler subformulas via a divide‑and‑conquer step that inspects second‑order derivatives of a lightweight “oracle” network, then polishes structure and constants with brief MCTS/GP (Monte Carlo tree search/genetic programming) refinement. The pipeline enforces dimensional consistency and favors compact, interpretable expressions, addressing common pitfalls of overfitting or unit-violating candidates. The authors plan to extend the framework to calculus-aware operators for PDE‑style laws and improve robustness to noisy lab data. PhyE2E highlights a neuro‑symbolic path toward automated, interpretable scientific discovery that can generalize beyond space physics.