Interactive Learning with Low‑Performance Data for Accurate CO2 Reduction Catalyst Prediction

Abstract Emerging machine learning (ML) approaches offer transformative strategies to accelerate discovery for challenging objectives, such as developing practical catalysts through rational design. We propose an iterative learning framework and strategy that effectively enhances the predictive capability of the ML model, thereby accelerating catalyst development. By leveraging low-performance catalytic results, we have proven a practical, scalable framework for data-efficient catalyst development and for the broader application of ML methods in heterogeneous catalysis. The ML approach effectively captures the intrinsic relationship between metal type/content and catalytic performance, thereby enhancing prediction accuracy through iterative optimization. The case example regards the development of metal-supported catalysts for CO2 reduction. The results show that incorporating low-performance catalytic results into the iterative learning process markedly improves the predictive accuracy of the random forest (RF) model (R2 increased from 0.80 to 0.86, and mean square error (RMSE) decreased from 1.20% to 0.89%). The deviations between the experimental and predicted values for the CO2 reduction performance of the four catalysts identified through this approach were all less than 2%, demonstrating the model's reliability and predictive capability. Among them, the 1.5Pt2Fe/TiO2-R catalyst exhibited high CO2 reduction activity and long-term stability (>200 h), highlighting the practical potential of ML-guided catalyst design. The use of low-performance catalytic data outperforms high-performance ones in improving predictive accuracy when coupled with a suitable iterative learning framework, providing mechanistic insight into how they help the model define the boundaries of the feasible performance region.