Flexible On-the-Fly Active Learning for Molecular Infrared Spectra

Abstract Infrared spectroscopy serves as an indispensable analytical tool across materials science, pharmaceutical development, and the life sciences. Computational insights derived from first-principles methods enhance the interpretation of vibrational spectra, yet their application remains hindered by high computational cost. Machine-learned interatomic potentials offer a route to accelerate these predictions by multiple orders of magnitude without sacrificing ab initio accuracy. Here, we present an on-the-fly active learning framework for the automated training of such potentials specifically tailored to infrared spectral prediction. Our approach employs a multitask acquisition strategy, in which model performance is evaluated iteratively on designated objectives, avoiding the dependence on fixed training sets. We demonstrate that this multitask scheme creates interatomic potentials that exhibit robustness and fidelity in complex workflows for calculating infrared spectra.