Extrapolation Techniques in Database Construction for Machine-Learning Potentials: Achieving Subchemical Accuracy in Sampling Conformal Funnels in Catalytic Processes

We present a computational workflow, the conformal sampling of catalytic processes enhanced with extrapolation techniques (CSCP-X), for constructing machine-learning interatomic potentials (MLIPs) to accelerate the exploration of first-principles potential energy surfaces for complex catalytic reactions. The MLIPs developed within the enhanced CSCP framework achieve subchemical accuracy error (∼0.03 eV) with respect to the energetics of the generating first-principles DFT approach, that is the targeted accuracy for predictive catalyst screening, while maintaining the same computational efficiency and throughput of the original CSCP method, still requiring only up to two iterative active learning steps and with a very modest increase in the size of the dataset. The CSCP-X approach is demonstrated on the methanol (CH₃OH) decomposition reaction pathway on (111) and (100) metal facets of eight metal catalysts (TM = Fe, Co, Ni, Cu, Pd, Ag, Au, and Pt). By combining mechanistic invariance with mathematical series extrapolation techniques, we achieve the ability to fully exploit the idea of "conformal funnels" in reaction processes. The generated MLIPs not only predict very accurately reaction energy barriers via nudged elastic band (NEB) transition state searches but also prove to be able to uncover alternative reaction paths and changes in the reaction mechanism with reduced energy barriers, reaching mechanistic transferability <i>en par</i> with the best physics-based models. CSCP-X offers itself as an operative, robust, and highly efficient pathway for accelerating the discovery of novel catalytic materials through high-throughput catalyst screening.