Active learning driven multi-property inverse design of spinel solar cells

Spinel materials offer promising potential for solar cell applications due to their highly tunable compositions and unique structures. However, developing efficient and stable spinel solar cells is hindered by data scarcity and the challenge of meeting multi-property requirements through traditional experimental and data-mining approaches. To address these limitations, we developed a comprehensive multi-property inverse design framework that integrates active learning to mitigate data scarcity and inverse design to efficiently identify optimal spinel compositions using the most informative data points. Within this framework, we applied a Multi-Task Gradient Boosting Machine model, which simultaneously predicts critical properties—bandgap values, bandgap type, and stability—achieving Area Under Curve scores of 0.86, 0.91, and 0.86, respectively. Leveraging this approach, we systematically explored over 10¹³ compositional combinations and identified 168 spinel materials that satisfy stringent criteria for high-performance solar cells. By employing interpretable machine learning techniques, we further analyzed the structure–property relationships of these materials, yielding actionable insights for targeted material design. This work not only accelerates the discovery of advanced spinel-based solar cells but also establishes a versatile strategy applicable to the design of other functional materials.