How unconstrained machine-learning models learn physical symmetries

A team of researchers from EPFL, led by Michelangelo Domina, has published a significant paper investigating how flexible, 'unconstrained' machine learning models learn the fundamental symmetries of physics. Traditionally, building AI for scientific simulation requires baking physical laws—like rotational equivariance—directly into the model's architecture using rigid, constrained mathematical forms. This new research challenges that paradigm by rigorously analyzing two popular unconstrained architectures: a transformer-based graph neural network for atomistic simulations and a PointNet-style model for particle physics. Using their novel diagnostic metrics, the team tracked how symmetry information propagates through the models' layers during training.

The key finding is that these highly expressive, unconstrained models can indeed learn approximate equivariant behavior to a high degree of accuracy, primarily through a simple strategy of data augmentation during training. This means that instead of being hardwired from the start, the models discover the underlying physical rules from the data itself. More importantly, the research establishes a formal framework for diagnosing 'spectral failure modes'—specific ways in which models break physical laws. This diagnostic power enables a hybrid approach: strategically injecting the minimum required inductive bias to guarantee physical fidelity, while preserving the model's flexibility and scalability. The work provides a blueprint for building more accurate and stable AI for materials science, chemistry, and physics without sacrificing the power of modern, general-purpose architectures.