Harnessing Active Learning with Machine Learning Potentials to Investigate Deactivation Mechanisms in Transition Metal-Doped Alumina Catalysts

Machine learning interatomic potentials (MLIPs) are increasingly developed and employed to study various catalytic processes, including surface reactivity, diffusion barriers, and reaction pathways. However, accurately predicting the behavior of complex phenomena that span multiple length and timescales, such as catalyst deactivation, remains a significant challenge. To overcome this limitation, we have trained MLIPs using first-principles DFT data on alumina doped with transition metals (TMs)—Fe, Co, Ni, Ru, Rh, and Pd—in cluster and atomic configurations to investigate deactivation mechanisms in these systems. Initially, we benchmarked two different MLIP architectures: MACE and NEP. Our results demonstrate that MACE significantly outperforms NEP in predicting energies and forces, with both the root mean squared error (RMSE) and mean absolute error (MAE) being more than an order of magnitude lower. To ensure reliable performance across diverse configurations, we implemented an active learning protocol, refining the model by focusing on challenging configurations. Using an uncertainty quantification metric, the MACE model was iteratively trained through NVT molecular dynamics (MD) simulations in LAMMPS, followed by DFT single-point energy (SPE) calculations on randomly sampled MD configurations. Configurations with uncertainties exceeding a threshold (0.3 eV/Å) were added to the initial training set, and the model was retrained until convergence. To streamline this iterative process, we developed an in-house automated software solution—Dynamic Refinement and Iterative Validation Exploration Software for Chemicals and Materials (DRIVES-CheM). Employing DRIVES-CheM to obtain the optimized MACE parameters, we performed metadynamics calculations to investigate catalytic deactivation mechanisms in doped alumina surfaces with different TMs. We observed significant differences in the energy barriers between different regions of TM-doped alumina surfaces. As such, this automated framework has significantly enhanced the efficiency of our training and evaluation workflow, providing a robust tool for managing complex computational tasks and extending our approach to explore other systems and processes responsible for catalytic deactivation.