Key points are not available for this paper at this time.
Foundation machine-learned interatomic potentials promise rapid access to high-quality potential energy surfaces, but their fitness for gas-phase chemical kinetics remains largely untested. Here we benchmark the Universal Models for Atoms (UMA) foundation model for gas-phase kinetics applications relevant to combustion and atmospheric chemistry using our automated KinBot workflow across 12 representative systems. We compare optimized structures and ZPE-corrected energetics against the parent level of theory, ωB97M-V/def2-TZVPD, and assess pathway discovery, stationary-point fidelity, and downstream kinetic inputs such as conformer ordering and one-dimensional hindered rotor scans. UMA reliably identifies the expected reaction channels over broad regions of chemical space, including pathways that are also a challenge for ab initio methods. Single-point DFT corrections at UMA geometries are an efficient way to improve energies. Overall, we suggest a practical hybrid workflow in which UMA performs inexpensive exploration and sampling, while DFT refinement is reserved for the important regions of the PES. Our results indicate that UMA can substantially accelerate rate coefficient calculations for gas-phase systems, motivating future work on uncertainty quantification, targeted finetuning, and Δ-learning corrections toward a gas-phase kinetics-specialized foundation model.
Kendall et al. (Mon,) studied this question.