Unveiling 3D hadron structure with interpretable and explainable AI methods

I will present recent results obtained by the EXCLAIM collaboration, establishing new connections between nuclear physics and computer science. Interpretable AI methods (Symbolic Regression) and explainable AI methods (variational autoencoders and diffusion models) are applied to hadron tomography. In the first case, we use SR’s uniquely human-readable machine learning (ML) architecture to construct a framework for physics-informed ML, a new paradigm of the relationship between machine learning and theory development, and harness it to perform

hypothesis testing and guided extrapolation on lattice QCD data; diffusion models are used to perform and en-to-end analysis of deeply virtual exclusive scattering data for a more direct extraction of quark and gluon angular momentum and other dynamic properties from data. This new paradigm pushes the field beyond black-box machine learning, whereby ML tools informed by physical principles can break longstanding bottlenecks to revolutionize the state of the art in extracting hadron structure.