<Graphical abstract>

Enhancing the accuracy of predictive models for catalytic performance is critical for effectively leveraging accumulated data and optimizing catalysts and operating conditions. This study evaluates the potential of hierarchical regression models in predicting the hydrogen formation rate during ammonia decomposition at elevated reaction temperatures. Using a literature database, we trained sequential, random, and inverse hierarchical regression models, alongside non-hierarchical models, on data collected at lower temperatures. A comparison of predictive accuracies revealed that the sequential hierarchical models, organized by increasing temperature, provided the highest accuracy. This finding illustrates the model’s capability to capture the nonlinear relationship between reaction temperature and catalytic performance. In contrast, both random and inverse hierarchical models performed worse than non-hierarchical models, highlighting the importance of subset training order on hierarchical model performance. Examining the effects of subset configuration, hierarchy level, and algorithm type on the predictive accuracy of sequential hierarchical models for ammonia decomposition provides insight into the optimized design of these models. The successful application of hierarchical regression models in the prediction of catalytic performance demonstrates their advantages in capturing the complexity inherent in the heterogeneous catalysis database. Ultimately, the hierarchical model shows strong potential for robust and generalizable predictions under various reaction temperatures in ammonia decomposition, extending beyond the original training domain.