Chemistry-wise transfer active learning accelerated investigation of high-entropy alloys towards CO₂ reduction

Machine learning technologies have been employed to explore enormous chemical space of high-entropy alloys (HEAs) for electrocatalytic applications. However, their performance awaits further improvement due to the limited datasets and complex interaction between adsorbates and HEAs. In this study, we present a chemistry-wise transfer active learning framework designed to address these challenges by systematically capturing chemical interaction across increasing levels of compositional complexity. Specifically, we employ a sequential transfer learning strategy that progressively learns adsorption process from unary and binary intermetallic systems, through ternary compounds, and ultimately combined with active learning to HEAs. This hierarchical approach significantly enhances the prediction accuracy of adsorption energy for the intermediates involved in CO₂ reduction, outperforming the models trained directly on the limited HEA dataset. Further analysis using t-distributed stochastic neighbor embedding (t-SNE) reveals that the improved performance arises from the effective transfer of chemically relevant features across domains. Our results demonstrate that chemistry-wise transfer active learning not only improves predictive capability in data-scarce regimes but also holds promise for accelerating the investigation of a broad range of functional materials with reduced cost and energy consumption.