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

Multireference perturbation theory methods, such as complete active space second-order perturbation theory (CASPT2), are often employed to recover the missing electron correlation from multiconfigurational zeroth-order wavefunctions. Here, we introduce the data-driven CASPT2 (DDCASPT2) method to capture dynamical electron correlation using features generated from lower-level electronic structure methods such as Hartree-Fock and complete active space self-consistent field (CASSCF) theory. We examine the effects of system size, basis set size, and number of two-electron excitations using a small, but diverse, set of molecules. We also provide insights into our physics-based feature set using SHapley Additive exPlanation (SHAP) analysis, a feature analysis method based on cooperative game theory. In this paper, we utilize these insights to introduce a DDCASPT2 method, which provides a machine learning-based alternative to traditional single- and multi-state CASPT2 for capturing dynamical electron correlation with near-CASPT2 quality accuracy.