Two-dimensional infrared spectroscopy of solute-solvent complexes from linear-scaling DFT and machine learning.

Two-dimensional infrared (2DIR) spectroscopy captures vibrational correlations on femtosecond timescales, offering direct insights into hydrogen(H)-bonding dynamics and other ultrafast molecular processes. However, interpreting these spectra requires simulations that accurately describe solute-solvent interactions over realistic timescales and system sizes. While classical approaches using empirical frequency maps are common, ab initio molecular dynamics (AIMD) offers a more rigorous alternative by treating dynamics and vibrational frequencies on a consistent theoretical level. The primary drawback of AIMD is its high computational cost, which typically limits simulations to short trajectories. Here, we introduce a hybrid strategy that combines linear-scaling density functional theory (LS-DFT) with a machine-learned (ML) interatomic potential. We use short LS-DFT simulations to generate reference energies, forces, and electron-density-derived dipole moments, which then serve as training data for a DeepMD model. The resulting ML potential allows nanosecond-scale dynamics at a fraction of the ab initio cost. We demonstrate this approach for N-methylacetamide in methanol, a model system known to form distinct H-bonded subpopulations. A key advantage of our method is that it bypasses the need for empirical frequency maps. Instead, molecular dipoles are learned directly from the electron density, and instantaneous vibrational frequencies are calculated from stable numerical Hessians. The resulting linear and 2DIR spectra show excellent agreement with experiment, accurately reproducing the characteristic doublet structure of the amide I band. This framework provides a practical and accurate route to simulating vibrational spectra at the AIMD level of theory for a wide range of IR-active solutes and H-bonded complexes.