Learning Full Configuration Interaction Electron Correlations with Deep Learning

Coupled cluster theory on modern heterogeneous supercomputers

“…Many of the reported works on molecular applications of ML models propose that these models are not just proxies but computational ends for theoretical molecular quantum chemistry methods. Nonetheless, many of these methods rely on correlations between families of molecules, where basic parameters such as the molecular topology, molecular local descriptors, and narrow electronic property classification of relationships are emphasized ( Haghighatlari and Hachmann, 2019 ; Corzo et al, 2021 ; Fung et al, 2021 ; Keith et al, 2021 ; Unke et al, 2021 ). As a consequence, the use of current ML models as an alternative for ab-initio quantum-chemical techniques may be less relevant when elucidating molecular phenomena for which unprecedented electron correlation effects and possibly the interactions between electronic and nuclear degrees of freedom play a fundamental role, e.g., redox processes, chemical reactions, photochemical processes, organometallic catalysis, etc.…”