Unlocking Potential of Pyrochlore in Energy Systems via Soft Voting Ensemble Learning

Abstract In traditional machine learning (ML)‐based material design, the defects of low prediction accuracy, overfitting and low generalization ability are mainly caused by the training of a single ML model. Here, a Soft Voting Ensemble Learning (SVEL) approach is proposed to solve the above issues by integrating multiple ML models in the same scene, thus pursuing more stable and reliable prediction. As a case study, SVEL is applied to develop the broad chemical space of novel pyrochlore electrocatalysts with the molecular formula of A ₂ B ₂ O ₇ , to explore promising pyrochlore oxides and accelerate predictions of unknown pyrochlore in the periodic table. The model successfully established the structure‐property relationship of pyrochlore, and selected six cost‐effective pyrochlore from the periodic table with a high prediction accuracy of 91.7%, all of which showed good electrocatalytic performance. SVEL not only effectively avoids the high costs of experimentation and lengthy computations, but also addresses biases arising from data scarcity in single models. Furthermore, it has significantly reduced the research cycle of pyrochlore by ≈ 22 years, offering broad prospects for accelerating the development of materials genomics. SVEL method is intended to integrate multiple AI models to provide broader model training clues for the AI material design community.