Machine-learning to predict anharmonic frequencies: a study of models and transferability.

Area of Science:

• Computational Chemistry
• Physical Chemistry
• Chemical Physics

Background:

• Accurate thermochemical properties require anharmonic effects.
• Describing anharmonicity is crucial in physical and chemical physics.
• Electronic structure methods are increasingly accurate, necessitating advanced post-processing.

Purpose of the Study:

• To calculate anharmonic frequencies of hydrogen-halides and halogenated hydrocarbons.
• To develop predictive models for anharmonic frequencies using machine learning.
• To assess the transferability of these models to new chemical systems.

Main Methods:

• Calculation of anharmonic frequencies using normal mode analysis and vibrational self-consistent field (VSCF).
• Development of predictive models via multilinear regression and gradient boosting regression.
• Utilizing harmonic model-based descriptors for regression analysis.

Main Results:

• Gradient boosting regression accurately predicts anharmonic frequencies.
• Multilinear regression provides reasonable predictions, capturing mode-to-mode couplings.
• Machine-learned models demonstrate transferability to larger, unseen molecules.

Conclusions:

• Machine learning models, particularly gradient boosting, can reliably predict anharmonic frequencies.
• Simple regression models offer valuable insights into anharmonic behavior.
• The developed models show excellent potential for application to diverse chemical systems.