Tensorial Properties via the Neuroevolution Potential Framework: Fast Simulation of Infrared and Raman Spectra

Area of Science:

• Computational chemistry and materials science.
• Spectroscopy and vibrational dynamics.
• Machine learning applications in physical sciences.

Background:

• Infrared and Raman spectroscopy provide crucial insights into molecular and material dynamics.
• Atomic-scale simulations for spectral prediction are often computationally expensive or rely on limiting approximations.
• Existing machine learning (ML) methods for spectral prediction have limitations in efficiency and applicability.

Purpose of the Study:

• To develop a novel ML-accelerated approach for accurate and efficient prediction of infrared and Raman spectra.
• To generalize the neuroevolution potential concept for predicting tensorial properties.
• To provide a readily accessible computational tool for the scientific community.

Main Methods:

• Introduction of the tensorial neuroevolution potential (TNEP) scheme, generalizing neuroevolution potentials to predict rank one and two tensors.
• Development of ML models for dipole moment, polarizability, and susceptibility using TNEP.
• Application of TNEP to predict spectra for liquid water, PTAF-, and BaZrO3.

Main Results:

• The TNEP approach demonstrates superior accuracy and computational efficiency compared to existing ML models.
• Successful prediction of infrared and Raman spectra for diverse systems including liquids and solids with anharmonicity.
• Validated TNEP models for molecular dipole moment, polarizability, and susceptibility.

Conclusions:

• The TNEP scheme significantly enhances the efficiency and accuracy of ML-based spectral prediction.
• This methodology overcomes limitations of traditional simulations and prior ML approaches.
• The open-source implementation (gpumd) facilitates widespread adoption and advancement in computational spectroscopy.