Simulation of Raman-Spectra of water using machine learning potentials

In this paper, we present an alternative method to ab-initio molecular dynamics (AIMD) simulations for Raman spectra calculations of water molecules which can be computationally expensive. We offer a more efficient method for spectra calculation by utilizing neural network potential (NNP) to reduce computational costs while maintaining accuracy comparable to AIMD simulations. The Deep Polar (DeepPol) model, trained using data from density functional theory simulations, predicts polarizabilities without relying on central atom assignments, allowing for environment-dependent contributions from all atoms. We validate the simulated spectra by comparing results to both AIMD simulations and experimental Raman spectra, analyzing the temperature dependence of the OH stretching band. Key parameters such as sampling time, correlation depth, and system size are systematically investigated to understand their effects on spectral outcomes. The findings demonstrate that machine learning potentials, when integrated with molecular dynamics simulations, provide a computationally efficient framework for simulating Raman spectra, with potential applications beyond water systems. • Neural network potentials simulate Raman spectra with accuracy near ab-initio models. • Generalized DeepPolar model predicts polarizabilities without central-atom assignments. • Simulated spectra validated against ab-initio and experimental Raman data. • Study examines how various factors affect the simulated Raman spectra.