Enhancing the Efficiency and Flexibility of AutoMeKin: Integrating ORCA and Machine-Learning Potentials

Abstract Automated reaction discovery is often limited by the cost of high-level refinement of large reaction networks. Here, we extend AutoMeKin, an automated reaction-discovery program, by presenting a complete integration of ORCA 6 into its high-level refinement workflow, providing access to a broad range of electronic-structure methods through a package freely available to academic users. We also integrate the UMA and MACE-OMOL-0 machine-learning interatomic potentials and Δ-ML energy corrections, enabling automated refinement of reaction networks using conventional electronic-structure methods, foundation ML potentials, or ML-corrected quantum-mechanical approaches within a unified framework. The workflow is assessed for five systems, monoethanolamine, three cyclobutadiene-derived radicals, and N-[bis(2H-triazol-4-yl)methylidene]hydroxylamine, spanning both in-distribution and challenging out-of-distribution chemistry. Comparisons with reference data show that UMA and MACE-OMOL-0 generally recover the relevant reaction mechanisms and provide energy profiles of comparable quality to DFT at substantially reduced cost. RRKM analyses further reveal that vibrational frequencies can significantly affect product population fractions, especially for loose transition states. These results show that the new implementation increases the accessibility and flexibility of AutoMeKin and further support the view that foundation ML potentials can serve as practical surrogates for DFT in automated reaction-network discovery.