Benchmarking the UMA Foundation Interatomic Potential for Gas-Phase Chemical Kinetics

Area of Science:

• Computational Chemistry
• Chemical Kinetics
• Machine Learning

Background:

• Machine-learned interatomic potentials offer rapid access to potential energy surfaces.
• Their application to gas-phase chemical kinetics, crucial for combustion and atmospheric chemistry, is largely unexplored.

Purpose of the Study:

• To benchmark the Universal Models for Atoms (UMA) foundation model for gas-phase kinetics.
• To evaluate UMA's performance in reaction pathway discovery and kinetic data generation using the KinBot workflow.

Main Methods:

• Benchmarking UMA across 12 diverse gas-phase systems using the automated KinBot workflow.
• Comparing UMA-optimized structures and energies against high-level ab initio theory (ωB97M-V/def2-TZVPD).
• Assessing stationary-point fidelity, conformer ordering, and hindered rotor scans.

Main Results:

• UMA reliably identifies relevant reaction channels, even those challenging for traditional ab initio methods.
• A hybrid workflow combining UMA exploration with DFT refinement provides efficient and accurate energy calculations.
• UMA significantly accelerates the computation of rate coefficients for gas-phase systems.

Conclusions:

• UMA shows strong potential for accelerating gas-phase kinetics studies.
• A hybrid UMA-DFT workflow is practical for exploring potential energy surfaces and refining key kinetic parameters.
• Future work should focus on uncertainty quantification and specialized fine-tuning for kinetics applications.