Assessing the quality of machine learned force fields for dynamical properties using quasi-elastic neutron scattering

The richness of the insight provided from the combined use of simulation and neutron spectroscopy has been extensively proven in recent times. Molecular Dynamics has the capability of studying the same length and time scales as that of Quasi-elastic Neutron Scattering (QENS), however the most accurate descriptions of intermolecular forces provided by ab initio methods are unfortunately severely limited by current computational power. The recent emergence of Machine Learning Force fields presents the possibility of retaining elements of an ab initio level description, but allowing the simulation of nano-scale dynamics required for the comparison with experimental QENS results. Early results show that these new models are capable of correctly predicting structural properties, however as yet there have not been robust tests on the accuracy of their dynamics. We propose to perform the first tests of these models for their nano-scale dynamics to assess whether these models could represent a paradigm shift in how we analyse QENS data.