Machine learning potentials accurately reproduce vibrational dynamics in complex environments.

Vibrational spectroscopy provides a bond-specific view of molecular structure and dynamics, but translating spectroscopic observables to local environments requires the development of accurate models to compute spectroscopic observables from simulations. Recent advances in machine learning interatomic potentials (MLIPs) provide ab-initio-like accuracy at low computational cost, opening new opportunities for developing general, transferable models that can predict spectroscopic observables without system-specific parameterization. Here, we benchmark the performance of the Universal Model of Atoms (UMA), a recently developed MLIP, for predicting IR absorption spectra and picosecond frequency fluctuations of an ester carbonyl in a range of solvents. UMA results are compared with two established approaches: an empirical frequency map parameterized for the ester carbonyl and the semiempirical tight-binding method, GFN2-xTB. We find that UMA reproduces experimental observables with accuracy comparable to traditional methods, while offering broader generality and efficiency.