Applying machine learning (ML) to assist the interpretation of analytical spectra associated with complex product mixtures and reaction outcomes presents distinct challenges with respect to data scarcity, featurization, and contextual metadata. Nevertheless, recent developments in automated reaction prediction present compelling opportunities for semi-automated approaches to build diverse reactivity training datasets that can be used for spectral interpretation. In this talk we’ll describe our recent development of an automated reaction prediction framework, Yet Another Reaction Program (YARP), that addresses several challenges of characterizing deep reaction networks using computational methods. Compared with prevailing methods, YARP near perfectly reproduces established reaction pathways and products, without the use of domain knowledge. In addition, YARP also discovers new kinetically relevant reaction pathways and exhibits nearly 100-fold reductions in computational costs. In turn, this combination of ultra-low cost and high reaction-coverage is being applied to generate training reaction network data for use in ML-based reaction prediction models. Time permitting, we will discuss how such models can be used to learn the reactivity relationships generated by physics-based methods like YARP, to semi-automatically assist the interpretation of analytical spectra associated with complex reaction mixtures.

How Things Fall Apart: Using Automated Reaction Prediction as a Tool for Characterizing Complex Reaction Outcomes

ACS Fall 2021

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