Multiscale Machine Learning Prediction of Infrared Spectra of Solvated Molecules.

Abstract

We introduce a multiscale machine-learning molecular dynamics (MD) strategy for simulating infrared spectra of solvated molecules. Our approach integrates an efficient sampling of environmental configurations with a hierarchical model that predicts forces and dipole moments as analytical derivatives of the energy, allowing IR spectra simulations from MD trajectories. Solvent effects are incorporated through a molecular mechanics (MM) representation of the environment embedded within the ML description of the solute. Applied to representative biorelated systems, the resulting ML/MM framework reproduces experimental spectra with high fidelity and accurately captures solvent-driven vibrational shifts. This approach provides a computationally efficient and robust route for describing solvent effects in vibrational spectroscopy.