Deep Potential-Driven Molecular Dynamics of CO Ice Analogues: Investigating Desorption Following Vibrational Excitation

Abstract

We present a new deep learning-based machine learning potential (MLP) for molecular dynamics simulations of solid carbon monoxide (CO), capable of accurately describing CO vibrations both in the fundamental state and in highly excited vibrational states, up to approximately v = 40. The MLP is based on the combination of high-dimensional neural network atomic potentials using the DeePMD-kit package, trained on prior ab initio molecular dynamics (AIMD) data, with selective treatment of the excited molecule allowing us to capture complex energy redistribution dynamics in condensed-phase environments. In particular, the MLP is capable of accurately describing the desorption process of a single CO molecule within an aggregate of 50 CO molecules, in excellent agreement with both previous theoretical predictions and experimental measurements. The MLP provides a much finer description of the translational and rotational energy distributions, capturing their character with high fidelity and allowing a more detailed comparison with experimental results. Furthermore, the analysis of the rotational energy, resolved over specific translational energies, revealed new insights into the coupling between translational and rotational degrees of freedom during the photodesorption process. This novel approach opens new perspectives for extensive statistical studies on desorption energies and detailed investigations of surface molecule excitations and the exploration of larger-scale models incorporating periodic boundary conditions to simulate more realistic CO aggregates.