ICML 2026 Workshop on AI for Physics

We will explore how well today’s LLMs and autonomous agents can generate physically consistent, rigorously testable hypotheses; derive predictions with correct assumptions, units, and constraints; and interpret outcomes from simulations or experiments. The workshop will discuss common failure modes in physics reasoning and enabling directions such as tool-augmented agents, retrieval over scientific papers and code, and structured memory for multi-step derivations and verification.

This topic focuses on learning-based simulators and emulators for complex physical processes, including PDE-governed dynamics, turbulence, plasma systems, cosmology, and detector-level simulation. We highlight neural operators, physics-constrained generative models, differentiable simulators, hybrid solvers, and error-controlled emulation strategies in challenging regimes such as long-horizon rollouts, stiff dynamics, multiscale coupling, and rare events.

We focus on recovering physical parameters, fields, or latent states from indirect measurements through forward models, often involving simulators. The workshop will cover likelihood-free and simulation-based inference, differentiable and amortized inference, inverse design, and methods for handling nuisance parameters, calibration errors, selection effects, and simulation-measurement mismatch.

This topic studies learned, stateful models of physical systems that map multimodal observations to compact latent representations and predict dynamics under partial observability. We emphasize extrapolation to unseen regimes, transfer from simulation to experiment, and hybrid approaches that combine physical structure with data-driven representations, with evaluation centered on out-of-distribution reliability.

We highlight physics areas that lack large, standardized, and openly accessible resources, and aim to catalyze community efforts in dataset creation, benchmark design, and reproducible evaluation. We will discuss synthetic data with controllable realism, weak supervision, targeted data acquisition, active learning, adaptive measurement, and autonomous experimentation.