Uncertainty Sampling-Based Efficient Data Generation for Development of Machine Learning Model to Predict Catalyst Degradation

Assessment of the long-term stability of a laboratory-developed catalyst is an essential step to determine if the catalyst can be commercialized. Currently, long-term stability testing requires long running time on stream studies which are costly and generally not performed in academic settings. Accelerating the ability to assess the long-term stability is a critical step to improve rates of commercialization for novel laboratory-developed catalysts. The reverse water-gas shift (RWGS) reaction is a critical step in the Power-to-liquids (PtL) technology, which uses renewable electricity to provide sustainable liquid fuels, and is a component unit operation in the strategy to decarbonize the production of fuels and chemicals. While many promising catalysts have been reported in the literature, these catalysts require high operating temperatures in order to be thermodynamically favorable to the desired products, raising questions as to the long-term stability of the catalyst materials under operating conditions.<br/>In this contribution, we have developed a machine learning (ML) model to rapidly predict the degradation of oxide-supported metal catalysts, such as Rh/rutile-TiO₂, of which single metal atom morphology has been reported to catalyze RWGS pathway with high selectivity [1]. As the generation of reactivity test data is highly costly, we adopted uncertainty sampling (US) approach, one of active learning (AL) techniques, to efficiently sample the experimental data. Using those experimental data collected by AL, several ML models were trained, and the best model was chosen to use for rapid prediction of catalyst degradation. In addition to the prediction of degradation by a regression model, useful knowledge on the relationship between experimental features and catalyst degradation was extracted by interpretable ML method such as SHAP (Shapley Additive exPlanations). This work is expected to give a guidance for data-efficient exploration of search space to map the relationship between experimental catalytic features and a catalytic property of interest.<br/><br/>[1] <i>J. Am. Chem. Soc. </i><b>2015</b>, 137, 3076-3084.