Toward Autonomous Scattering Experiments with Surrogate Models and Agentic AI

Machine-learning force fields (MLFFs) and surrogate models have rapidly advanced the ability to approximate atomistic energies, forces, and scattering-relevant observables at significantly reduced computational cost, raising new opportunities for using atomistic modeling to help guide experiments. In this talk, I will present two recent efforts that explore how surrogate physics models could be embedded within agentic AI frameworks for autonomous scattering experiments. First, I will introduce an agentic AI system that autonomously plans and executes a synchrotron X-ray diffraction experiment by integrating large language model-based reasoning with structured tool use. Second, I will discuss a machine learning-augmented Bayesian optimal experimental design approach for inelastic neutron scattering, in which learned surrogate models are used to accelerate posterior evaluation and guide data acquisition toward maximally informative measurements. Together, these examples illustrate how AI agents and surrogate models can enable closed-loop, autonomous scientific workflows. This talk points toward a future in which surrogate models, such as MLFFs, and experimentation are tightly integrated through agentic AI systems to enable more efficient and adaptive scientific discovery.