(121g) Going Right to the Top: High-Throughput Transition State Identification

Discerning reaction rates and their underlying chemical behavior, for example, regioselectivity, a priori to experimental investigation can be pursued by molecular modeling of transition pathways. The utility of a molecular modeling-based approach, such as those using electronic structure calculations, has a well-established history of success, albeit most commonly for a small collection of small systems. In this talk, I will describe modern data-driven strategies to overcome the debilitating computational demands of relying on quantum mechanical (QM) calculations as the workhorse for transition state identification. First, AIMNet2-rxn will be introduced, which is a machine learned interatomic potential specifically targeting generalized mechanistic modeling of chemical reactions with near QM calculation accuracy. By leveraging inexpensive AIMNet2-rxn atomic force predictions, minimum energy pathway search can readily be performed on a scale of millions of reactions per day. Importantly, this throughput is achieved on consumer grade graphics processing units (GPUs), representing critical first steps toward democratization and routine use of high-throughput mechanistic analysis. Furthermore, I will demonstrate the AIMNet2-rxn inferred Hessians can drive efficient transition state optimization within an accuracy threshold of 0.1 Å root-mean-squared deviation and 1.0 kcal mol^-1 barrier heights. While this leap in capabilities can be transformative, identifying suitable reactant-product geometry pairs is a prerequisite, which is nontrivial to achieve for a diversity of chemical science challenges. In the second half of my talk, I will discuss data-driven modeling strategies to bypass this limitation by directly generating transition state geometries from the prototypical framework of considering chemical reactions as bond rearrangements, thus eliminating the need to specify reactant and product geometries. Using an exhaustive training set of ~10⁵ reactions, it will be shown that transition state geometries can be robustly obtained with accuracy near that of the AIMNet2-rxn optimization using recent generative diffusion modeling strategies. Overall, this combination of data-driven interatomic potentials and direct data-driven molecular modeling presents itself as a powerful foundation for accelerated evaluation of chemical transition pathways.