Predicting inelastic neutron scattering spectra from the crystal structure via data-driven approach

Inelastic neutron scattering (INS) spectroscopy is an ideal tool to measure lattice dynamics, such as phonon dispersion, density of states, and anharmonicity. Due to the complexity of the INS spectroscopy, insight is often obtained by comparing the measured INS spectra with simulations on various structural models. Although it is possible to compute the phonon properties with modern electronic structure methods, such as density functional theory, the high computational expense is a significant bottleneck for its application. Over the past decade, machine learning has become a favored alternative to quantum mechanical calculations due to its high efficiency and versatility. In this study, we developed a quantum-based INS database and employed machine learning models to predict the force constants of any given materials and, subsequently, the phonon dispersion, density of states, and the corresponding INS spectra. This work offers a convenient and efficient approach for INS experiment planning and data analysis. It will also shed light on fast and accurate calculations of other properties related to lattice dynamics, such as heat capacity.