Accurate and Rapid Vibrational Spectra of Molecular Crystals from Machine-Learned Interatomic Potentials

Currently, vibrational spectra of molecular crystals often require computational methods to aid in band assignment, especially in the low-frequency region where no functional group specific transitions exist. DFT methods are the gold-standard for computational spectroscopy, but their high computational cost limits simulations to systems characterized by small unit cells. Universal machine-learned interatomic potentials (MLIPs) offer a much cheaper alternative, but their reliability for phonons in molecular crystals has been little explored. In this work, eleven MLIPs from the ANI, MACE or UMA families of models are tested on geometry optimizations, vibrational frequencies and normal mode character with the same computational workflow utilized by DFT. The resulting phonons are then combined with DFT-calculated IR and Raman intensities to obtain vibrational spectra through a hybrid ML/DFT approach. The method is tested on three dispersion-bound molecular crystals, BTBT, naphthalene and anthracene, and three hydrogen-bonded systems, β-glycine and the two polymorphs of the ethosuximide (ETS) drug. MACE-MH omat, UMA omat and UMA omc emerge as the best-performing models and reproduce the experimental spectra with fidelity comparable to DFT.