CASNet: Learning Complete Active Space Orbitals using Message Passing Neural Networks

Machine learning has demonstrated success in predicting molecular orbitals obtained from common single-configurational quantum chemistry methods, such as Hartree-Fock or Kohn-Sham Density Functional Theory. In this work, we present an extension of this supervised learning framework to multi-configurational quantum chemistry methods and compare different approaches for learning excited-state molecular wavefunctions. More specifically, by utilizing recent advances in message passing neural networks designed for learning molecular properties, we investigate the learning of molecular orbitals from State-Averaged Complete Active Space Self Consistent Field as a means of speeding up the corresponding calculations. We demonstrate the advantage of this general approach, referred to as \textit{CASNet}, over traditional orbital initialization techniques on different datasets of pentafulvene and evaluate its practical utility.