Δ-Machine Learning of Polarizability Tensors Using a Dipole Interaction Model.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning Applications

Background:

• Molecular polarizability is crucial for understanding light-matter and intermolecular interactions.
• Accurate and efficient methods for calculating polarizability tensors are essential.

Purpose of the Study:

• Introduce a novel model, Delta_PIM_CCSD, combining a polarizable dipole interaction model (PIM) with Delta-machine learning.
• Predict polarizability tensors with high accuracy and efficiency.

Main Methods:

• Developed Delta_PIM_CCSD model using PIM and Delta-machine learning.
• Adapted reference geometry for rotational symmetry by diagonalizing PIM polarizability tensor.
• Parameterized model to coupled cluster singles and doubles (CCSD) polarizabilities from QM7b dataset.

Main Results:

• Delta_PIM_CCSD achieved accuracy comparable to DFT/B3LYP at lower computational cost for QM7b-like molecules.
• Accuracy was maintained for QM9 dataset after basis set correction.
• Performance decreased for molecules smaller and more chemically diverse than the training set.

Conclusions:

• The combination of PIM and Delta-machine learning offers a promising approach for predicting polarizability tensors.
• Larger, more diverse datasets with high-level theoretical polarizabilities are needed for broader applicability.
• Incorporating atom-specific polarizabilities could further improve model performance.