Context, Culture, and Craft in Effective AI for Physics

Physics is often described as the science of first principles, yet in practice it is a patchwork of methods: designing instruments, running simulations, analyzing data, and building models, that each demand judgment, approximation, and craft. Artificial intelligence now threads through all of these activities, from controlling gravitational-wave detectors to accelerating materials discovery. But the goal is not to automate physics, it is to extend how we reason about it.

In this talk, I will explore how embedding physical structure into AI systems, particularly through Euclidean neural networks, allows models to respect the symmetries that govern real systems while remaining flexible enough to learn from data. These architectures, rooted in group representation theory, have transformed how we model atomic systems and are beginning to inform new ways of representing scientific knowledge itself. Drawing from examples across scales, from particles to materials to giant instruments, I’ll discuss how physics’ long-standing habits of approximation, symmetry, and modeling can guide the design of more interpretable and reliable AI, and how AI, in turn, may help us uncover new abstractions and bridge the experimental, computational, and theoretical sides of science.