Exploring Data-Driven Quantum Chemistry: From Chemical Compound Space to Quantum Chemical Calculations

Computational chemistry is a large set of methods that can be utilized to understand chemical phenomena, such as reactions, and can provide insights into the electronic properties of atoms and molecules. One family of methods that is regularly applied in these studies is molecular electronic structure theory—which refers to the application of quantum mechanics to molecular systems. Despite the success of electronic structure theory methods, such as density functional theory (DFT) and coupled-cluster, these calculations can become computationally prohibitive with respect to system size or when applied for large-scale computational screenings of databases. To this end, novel algorithms, based on machine learning (ML), have been introduced to accelerate quantum chemical calculations and for the exploration of chemical space. In this work, novel tools based on ML and topological data analysis (TDA) are introduced for the exploration of transition metal molecular catalysis utilized for C-H bond activation, acceleration of single reference methods, such as coupled-cluster singles and doubles (CCSD), and multi-reference methods, such as the variational two-electron reduced density matrix complete active space self-consistent field method (v2RDM-CASSCF) and the complete active space second-order perturbation theory (CASPT2). With respect to transition metal catalysis, a novel workflow based on DFT, TDA, and ML has been introduced for the exploration of the Fe(IV)-oxo species chemical space for C-H bond activation, using methane as a case study. Next, we introduce a novel graph neural network architecture, called PairGraphNet, to accelerate the elucidation of the CCSD correlation energy and apply methods from TDA to explore the topology of electron correlation. We also discuss the refinement of energies from v2RDM-CASSCF and utilize game theory to provide insights into the feature set. Lastly, we discuss the development a machine learning scheme to accelerate the recovery of the CASPT2 correlation energy. In this work, we introduce ML and TDA as tools for exploring large swaths of chemical compound space and accelerate quantum chemical calculations.