Machine learning (ML) approaches have been used for "over-the-arrow" reaction optimization by predicting experimental conditions, reagents, and catalysts that maximize the output (e.g., yield, selectivity) for the desired transformation. These data-driven models are generally data-hungry and require large amounts of closely related reaction data. Quantum chemical models are used to predict trends in reactivity and selectivity from first principles, often using transition state modeling. However, such approaches are time-consuming and not amenable to high-throughput screening. In this talk, I describe a hybrid approach using predictive statistical models trained against data from quantum chemical calculations. This approach enables quantitative structure-reactivity relationships for complex catalytic reactions for which conventional ML or QM methods either fail or are not amenable.<br/><br/>In the Buchwald-Hartwig amination, an industrially significant C-N bond-forming reaction, the rate-determining step varies according to the substrate used. This poses a considerable challenge in predicting reaction performance in a rapid and reliable fashion. Herein, we explore the incorporation of distinct mechanistic steps in the cross-coupling catalytic cycle into a predictive machine learning approach. QM data is used to train surrogate ML functions, which in turn are used to assess the reaction feasibility for a vast library of amine, aryl, and ligand substrates.

Incorporating mechanistic multi-steps in machine learning for optimizing catalytic reactions

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.