High-Pressure Inelastic Neutron Spectroscopy: Experimental Validation of Machine-Learned Interatomic Potential Energy Landscapes

Abstract

Validation of Machine-Learned Interatomic Potential Energy Landscapes Authors: Jeff Armstrong, Adam Jackson, Alin Elena Machine-learned interatomic potentials (MLIPs) promise near density-functional theory accuracy at a fraction of the computational cost, offering a route toward predictive atomistic modeling of molecular and condensed-phase materials. Yet their reliability beyond the training regime remains difficult to establish experimentally. Here we use pressure-dependent broadband inelastic neutron spectroscopy (INS) as an experimental probe of MLIP transferability. Using a low-background NiCrAl high-pressure clamp cell, we measure INS spectra of crystalline 2,5-diiodothiophene at 10 K under atmospheric pressure and at 1.5 GPa. MACE-based MLIPs, fine-tuned on targeted DFT data, reproduce the experimental spectra across 0-1200 cm-1 at both pressures and remain stable in finite-temperature molecular dynamics simulations at 300 K. The models capture systematic pressure-induced blue shifts arising from steric stiffening and reproduce an anomalous red shift near 453 cm-1 associated with pressure-modified intermolecular interactions. These results demonstrate that pressure-dependent INS can validate how an MLIP responds to a controlled thermodynamic perturbation, testing not only equilibrium structure but also the pressure-dependent curvature of the potential-energy surface. High-pressure INS therefore provides a practical experimental route for validating transferable machine-learned potentials for molecular materials.