Capturing Electron Correlation with Machine Learning through a Data-Driven CASPT2 Framework.

Area of Science:

• Quantum Chemistry
• Computational Physics
• Machine Learning

Background:

• Multireference perturbation theory, including complete active space second-order perturbation theory (CASPT2), is crucial for accounting for electron correlation in electronic structure calculations.
• Accurately describing electron correlation is essential for predicting molecular properties.

Purpose of the Study:

• To introduce a novel data-driven CASPT2 (DDCASPT2) method for capturing dynamical electron correlation.
• To assess the performance of DDCASPT2 across various system and basis set sizes.
• To leverage machine learning and SHAP analysis for improved accuracy in electronic structure calculations.

Main Methods:

• Development of the data-driven CASPT2 (DDCASPT2) method utilizing features from lower-level electronic structure theories (Hartree-Fock, CASSCF).
• Systematic examination of DDCASPT2 performance with varying system sizes, basis sets, and numbers of two-electron excitations.
• Application of SHapley Additive exPlanation (SHAP) analysis to interpret the physics-based features used in the DDCASPT2 model.

Main Results:

• The DDCASPT2 method effectively captures dynamical electron correlation.
• Performance was evaluated on a diverse set of molecules, demonstrating robustness with respect to system size and basis set variations.
• SHAP analysis provided insights into the physical relevance of the machine learning model's features.

Conclusions:

• The DDCASPT2 method presents a viable machine learning-based alternative to traditional CASPT2.
• Achieves accuracy comparable to conventional CASPT2 for dynamical electron correlation.
• Offers a promising direction for enhancing computational chemistry efficiency and accuracy.