Physicists

Neural surrogate models trained on physics simulation outputs are achieving faster-than-real-time inference with accuracy matching or exceeding traditional numerical methods across weather forecasting (GraphCast), materials discovery (GNoME), plasma physics (DeepMind/DIII-D), and cosmological parameter estimation. The 2023–2025 period saw these move from research demonstrations to production deployments at major physics facilities, with Nvidia's Modulus framework enabling widespread adoption. The trajectory is clear: for any simulation domain with sufficient training data, neural surrogates will eventually dominate, and physics subfields are generating that training data continuously.

The AI Scientist (Sakana AI, August 2024) and Agent Laboratory (2025) demonstrated autonomous systems that complete the full research cycle: generating hypotheses, designing and running experiments (in simulation), analyzing results, writing manuscripts, and performing peer review — with the AI Scientist producing papers that passed initial journal screening. These systems currently work in ML subfields where experiments run in software, but their architecture is explicitly designed to generalize. Physics subfields where 'experiments' are computational (condensed matter DFT, lattice QCD, computational fluid dynamics) are the direct next target, with timelines of 2–3 years for meaningful deployment.