(362f) Machine Learned Interatomic Potentials for Rapidly Exploring Organic Reactions

Formulating synthetic pathways that yield novel compounds and evaluating molecule or material stability are discovery steps found in nearly all chemical science fields. The common underpinning of these tasks is to profile chemical reactions, which oftentimes require the accumulated experience of expert chemists, time-consuming experimental characterization, or expensive computational resource investments. Chemical intuition and reaction heuristics have an undeniable track record; however, modern molecule and material discovery strategies, such as laboratory automation, demand a scalable solution for assessing thousands-to-millions of candidate pathways.

In this talk, I will describe RxnAIMNet: a machine learned interatomic potential (MLIP) for efficient large-scale reaction analysis with density functional theory (DFT) accuracy. We show that RxnAIMNet can perform minimum energy pathway searches, transition state optimization, intrinsic reaction coordinate analysis, and predict activation barriers within ~2 kcal/mol of DFT calculations. These reaction characterization tasks are carried out 10⁶ times faster than the reference quantum mechanical calculations, which enables 10⁵-10⁶ chemical pathways to be evaluated daily with modest computational resources. Despite our training dataset being constructed with simple bond breaking and forming rules, we observe that RxnAIMNet is transferable to prevalent synthesis mechanisms, such as Diels-Alder, biorthogonal click chemistry, and ring-opening metathesis. As a practical demonstration, we apply RxnAIMNet to construct a deep reaction network for β-D-glucose pyrolysis, a relevant species for biomass conversion, where kinetically relevant pathways can be traced to the major products identified experimentally. Overall, the RxnAIMNet model is well-positioned to address challenges at the frontiers of retrosynthetic analysis, deep reaction network construction, and computer-aided synthesis planning: a collection of tasks that are currently limited by a pervasive efficiency-accuracy trade-off.