Coupling Electronic Structure Theory with Machine Learning for Chemical Applications

Abstract: Our recent efforts on the development of new computational methods that couple quantum chemistry with machine learning will be discussed. First, a novel molecular fingerprinting method based on persistent homology, an applied branch of topology, that can encode the geometric and electronic structure of molecules for chemical applications will be presented. We have demonstrated its applicability on studies on non-covalent interactions between functional groups of materials and small molecules. The functional groups with enhanced CO2-philic groups can be introduced in the next generation of polymeric membranes with enhanced CO2/N2 separation performance. A short discussion on the applicability of the novel molecular fingerprinting method in catalysis and lanthanide separation will be given.

Second, we have developed a novel approach based on machine-learning algorithms and data that accelerates the convergence of computational chemistry methods. Our method uses quantum chemical data in order to learn correlated wave functions and provide highly accurate electronic energies with less computational effort. We have tested our data-driven method on the coupled-cluster singles-and-doubles (CCSD) level of theory. The data-driven CCSD (DDCCSD) is not an alchemical method since the actual iterative coupled-cluster equations are solved. DDCC provides a remarkable speed-up while it offers transferability from small molecules to larger molecular clusters.