Predictive Quantum Vibrational Spectra through Active Learning 4G-NNPs.

Area of Science:

• Condensed matter theory
• Computational chemistry
• Spectroscopy

Background:

• Predictive simulation of vibrational spectra in complex systems is computationally challenging.
• Modern condensed matter theory requires accurate modeling of nuclear quantum effects and anharmonicities.

Purpose of the Study:

• To develop and validate a novel computational framework for accurate vibrational spectral simulations.
• To integrate fourth-generation high-dimensional committee neural network potentials (4G-HDCNNPs) with path integral molecular dynamics.

Main Methods:

• Development of 4G-HDCNNPs using active learning and query-by-committee.
• Incorporation of nuclear quantum effects (NQEs), conformational entropy, and anharmonicities via path integral (PI) molecular dynamics.
• Seamless integration of nonlocal charge transfer effects with NQEs.

Main Results:

• Demonstrated accuracy in infrared spectral simulations for bulk water and air-water interfaces.
• Achieved accurate infrared spectra using predicted charges from 4G-HDCNNPs without explicit dipole moment training.
• Successfully integrated nonlocal charge transfer effects and NQEs.

Conclusions:

• The developed framework provides a simple, general, and practical paradigm for predictive spectral simulations.
• The methodology is free from empirical parametrizations and ad hoc fitting.
• Offers accurate modeling of complex condensed phases and interfaces.