Machine Learning the Complete Active-Space Wavefunction Reconstruction

Abstract

Machine-learned interatomic potentials enable multireference-quality reaction dynamics but do not retain the electronic structure underlying the predicted energies and forces. Recovering this information by independently performing complete active space self-consistent field (CASSCF) calculations is both expensive and potentially inconsistent. We introduce the Complete Active-Space Wave Function Reconstruction (CAS-WR) model, which learns the geometry-dependent electronic structure from CASSCF data and reconstructs the CAS wave function without repeated orbital optimization. Combined with uncertainty-based active learning, CAS-WR is demonstrated for the Diels--Alder reaction and \ce{TiC+}-catalyzed C--H activation. The approach enables trajectory-resolved analysis of multiconfigurational electronic structure across reactive ensembles. Preliminary applications to excited-state proton transfer and transition-metal photocatalysis suggest that CAS-WR can provide a general route toward recovering electronic structure along complex ground- and excited-state reactive dynamics.