Probing the Temporal Response of Liquid Water to a THz Pump Pulse Using Machine Learning-Accelerated Non-Equilibrium Molecular Dynamics

Abstract

Ultrafast, time-resolved spectroscopies enable the direct observation of non-equilibrium processes in condensed-phase systems and have revealed key insights into energy transport, hydrogen-bond dynamics, and vibrational coupling. While ab initio molecular dynamics (AIMD) provides accurate, atomistic resolution of such dynamics, it becomes prohibitively expensive for non-equilibrium processes that require many independent trajectories to capture the stochastic nature of excitation and relaxation. To address this, we implemented a machine learning potential that incorporates time-dependent electric fields in a perturbative fashion, retaining AIMD-level accuracy. Using this approach, we simulate the time-dependent response of liquid water to a 12.3 THz Gaussian pump pulse (1.3 ps width), generating 32 ns of total trajectory data. With access to ab initio-quality electronic structure, we compute absorption coefficients and frequency-dependent refractive indices before, during, and after the pulse. The simulations reproduce key experimental observables, including transient birefringence and relaxation times. We observe energy transfer from the excited librational modes into translational and intramolecular vibrations, accompanied by a transient, nonlinear response of the hydrogen-bond network and a characteristic 0.7 ps timescale associated with librational energy dissipation. These findings demonstrate the method's ability to capture essential non-equilibrium dynamics with theoretical time-dependent IR spectroscopy and establish a broadly applicable framework for studying field-driven processes in complex molecular systems.