Machine learning-assisted semiempirical quantum chemistry

• Applicant: Dr. Martin Stöhr
• Subject Area: Theoretical Chemistry: Electronic Structure, Dynamics, Simulation ; Physical Chemistry of Molecules, Liquids and Interfaces, Biophysical Chemistry ; Theoretical Chemistry: Molecules, Materials, Surfaces
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 534068594

Light-driven chemical processes play a central role in living organisms, from photosynthesis to vision, and are at the heart of several green energy solutions, from photovoltaics to efficient photocatalysts. Understanding and modeling such photochemical processes, however, usually requires computationally-expensive methods and non-adiabatic dynamics techniques able to capture chemistry across electronic states. As a result, available methods are typically limited to few atoms and model systems. This project aims to develop and apply a hybrid approach of semiempirical quantum chemistry (SE-QC) and machine learning (ML) in order to push the description and understanding of photochemical processes towards practically-relevant systems. The proposed adaptive ML-assisted SE-QC (AMASE-QC) method constructs a dynamic reduced-rank Hamiltonian using molecule-dependent parameters predicted via ML. First works with similar philosophy already showed very promising results of combined ML/SE-QC models for electronic ground states. In contrast to existing approaches we will, first, adopt an uncertainty-aware Delta-learning model, which seamlessly transitions to a fallback SEQC parametrization when outside the training domain of the ML model -- thus avoiding otherwise inevitable fatal generalization errors during dynamics. Second, we will incorporate our hybrid ML/SEQC model into a well-established post-Hartree-Fock machinery. In particular, we employ the resulting adaptive SE-QC method in the framework of floating occupation molecular orbitals (FOMO) and subsequent complete active space configuration interaction (CASCI). FOMO-CASCI has been shown to provide a reliable description across electronic states including potential degeneracies. Combined with non-adiabatic molecular dynamics techniques, it further enables efficient, yet accurate, photochemical simulations. Early work showed that this can also extend to the more cost-efficient variant of semiempirical FOMO-CASCI. This, however, requires tailor-made SE-QC parametrizations. The proposed AMASE-FOMO-CASCI formalism effectively uses exactly such tailor-made parametrizations in a fully self-contained, adaptive manner without the need of costly reference calculations and tedious reparametrizations for every new use case. With its lower computational costs, we envision AMASE-QC to push excited state and photochemical simulations beyond the length and time scales accessible to date. As part of this project, we will in particular study the role of realistic environments on the photodynamics of practically-relevant systems such as excited state proton transfer in photoactive proteins and the photoredox-catalytic properties of flavins and flavoproteins. Understanding how complex environments affect and guide photochemical processes thereby not only holds the key to unravel key photo-biological processes, but also provides design rules for next-generation artificial light harvesting and photocatalysts.

• DFG Programme: WBP Fellowship
• International Connection: USA
• Host: Professor Todd J. Martínez, Ph.D.