A foundational deep learning force field for simulating IR and Raman spectra via molecular dynamics

Accurate and efficient simulation of infrared (IR) and Raman spectra is essential for molecular identification and structural analysis. Conventional harmonic approaches neglect anharmonicity and nuclear quantum effects, while ab initio molecular dynamics (AIMD) remains computationally expensive. Here, we integrate a deep equivariant tensor attention network (DetaNet) with molecular dynamics to enable fast and accurate spectral simulations. The model is trained on the QMe14S dataset, comprising 186,102 small organic molecules with energies, forces, dipole moments, and polarizabilities, yielding a transferable machine learning force field. IR and Raman spectra are computed from time-correlation functions based on machine learning, molecular dynamics, and thermostated ring polymer molecular dynamics (TRPMD). Benchmark results for isolated molecules, including polycyclic aromatic hydrocarbons, show excellent agreement with experiments while achieving up to three orders of magnitude speedup over AIMD. The framework is further extended to molecular and inorganic crystals, aggregates, and polypeptides with minimal fine-tuning, maintaining high accuracy across diverse systems. This work establishes a foundational tensor-aware MLMD framework for efficient and accurate vibrational spectroscopy simulations.