Acquiring and transferring comprehensive catalyst knowledge through integrated high-throughput experimentation and automatic feature engineering.

Abstract

Solid catalyst development has traditionally relied on trial-and-error approaches, limiting the broader application of valuable insights across different catalyst families. To overcome this fragmentation, we introduce a framework that integrates high-throughput experimentation (HTE) and automatic feature engineering (AFE) with active learning to acquire comprehensive catalyst knowledge. The framework is demonstrated for oxidative coupling of methane (OCM), where active learning is continued until the machine learning model achieves robustness for each of the BaO-, CaO-, La₂O₃-, TiO₂-, and ZrO₂-supported catalysts, with 333 catalysts newly tested. The resulting models are utilized to extract catalyst design rules, revealing key synergistic combinations in high-performing catalysts. Moreover, we propose a method for transferring knowledge between supports, showing that features refined on one support can improve predictions on others. This framework advances the understanding of catalyst design and promotes reliable machine learning.