AI Can Predict, but Experiments Must Test: A New Role for Neutron Facilities in a Materials AI World

Abstract Materials AI is advancing at an exceptional pace and attracting growing investment, while the experimental scientific facilities and people needed to test and improve its predictions face mounting pressure. Unlike increasingly standardized data-center builds, neutron facilities are bespoke, source- and instrument-specific capabilities that take many years to build and mature, making lost capacity exceptionally difficult to replace. The response should be to embrace this shift by aligning experimental capability with the problems AI is increasingly being used to solve. For atomistic materials modeling, machine-learned interatomic potentials (MLIPs) provide a natural focus, and we propose coordinated experimental validation at two scales. Workhorse campaigns would combine automated room-temperature diffraction and low-temperature inelastic neutron scattering across reusable libraries of hundreds of well-characterized samples, with observables predicted before beamtime and compared systematically with experiment. Targeted parametric studies would examine selected systems across temperature, pressure, adsorption and guest loading using quasielastic neutron scattering, total scattering and complementary diffraction and spectroscopy. Together, these measurements would expose broad performance trends and physical origins of failure, while separating development, diagnostic and blind challenge data for subsequent model improvement. Neutron facilities are not alternatives to AI investment; they are experimental infrastructure through which atomistic materials AI can be tested against reality.