Multiscale Machine Learning Prediction of Infrared Spectra of Solvated Molecules.

Area of Science:

• Computational chemistry
• Spectroscopy
• Machine learning

Background:

• Simulating infrared (IR) spectra of solvated molecules is crucial for understanding molecular behavior.
• Traditional methods often face challenges in accurately and efficiently capturing solvent effects on vibrational spectra.

Purpose of the Study:

• To introduce a novel multiscale machine-learning molecular dynamics (ML/MM) strategy for simulating IR spectra.
• To accurately incorporate solvent effects into vibrational spectroscopy simulations.
• To provide a computationally efficient and robust method for analyzing molecular spectra.

Main Methods:

• Developed a multiscale machine-learning molecular dynamics (MD) strategy.
• Integrated efficient sampling of environmental configurations with a hierarchical ML model.
• Predicted forces and dipole moments as analytical derivatives of energy.
• Incorporated solvent effects using a molecular mechanics (MM) representation embedded within the ML description of the solute.

Main Results:

• The ML/MM framework reproduced experimental IR spectra with high fidelity for biorelated systems.
• Accurately captured solvent-driven vibrational shifts.
• Demonstrated computational efficiency and robustness in describing solvent effects.

Conclusions:

• The proposed ML/MM approach offers a powerful and efficient method for simulating IR spectra of solvated molecules.
• This strategy accurately accounts for solvent effects, crucial for understanding vibrational spectroscopy.
• The framework provides a robust computational route for spectroscopic analysis in complex environments.