Accurate molecular polarizabilities with coupled cluster theory and machine learning

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Molecular dipole polarizability is essential for describing molecular interactions and properties.
• Accurate prediction of polarizability is challenging due to its sensitivity to electronic structure.
• Existing methods like hybrid density functional theory (DFT) have limitations in accuracy.

Purpose of the Study:

• To develop a highly accurate and computationally efficient method for predicting molecular dipole polarizability.
• To benchmark the performance of a novel machine-learning approach against established quantum mechanical calculations and DFT.
• To provide insights into the deficiencies of DFT for polarizability prediction.

Main Methods:

• High-accuracy quantum mechanical calculations using linear response coupled cluster singles and doubles theory (LR-CCSD) for 7,211 small organic molecules.
• Development of a symmetry-adapted machine-learning model trained on LR-CCSD polarizability data.
• Validation of the machine-learning model on a diverse set of 52 larger organic molecules.

Main Results:

• The machine-learning model predicts LR-CCSD molecular polarizabilities with an error an order of magnitude smaller than hybrid DFT.
• The model demonstrates robustness and transferability, accurately predicting polarizabilities for larger, complex molecules.
• The atom-centered decomposition in the model highlights DFT's shortcomings in predicting this fundamental property.

Conclusions:

• Symmetry-adapted machine learning offers a powerful and accurate alternative for predicting molecular dipole polarizability.
• This approach significantly improves upon the accuracy of hybrid DFT methods at minimal computational cost.
• The findings pave the way for more reliable computational modeling of molecular interactions and properties.