Vision intelligence assists microstructural optimization of Ag-Bi-I perovskite-inspired materials

The slow progress in microstructural design for thin-film semiconductors remains a significant bottleneck in transitioning emerging materials into practical applications. While machine learning (ML) shows great promise in accelerating data analysis, effectively linking microstructural insights to actionable synthetic plans remains a complex challenge that surpasses the capabilities of a single model. We present the Daisy Visual Intelligence Framework, a computational platform using compound ML models for semiconductor microstructure optimization. Daisy integrates goal-oriented computer vision with an optimization agent to create a feedback loop between microstructural characterization and synthesis planning. Trained on historical synthesis and scanning electron microscopy (SEM) data for Ag-Bi-I compounds, an emerging family of perovskite-inspired materials with promise for indoor photovoltaics, the framework demonstrates the ability to learn effectively from imbalanced and heterogeneous experimental datasets. The artificial intelligence agent within Daisy identified new processing windows, which were experimentally validated to produce thin films with larger grain sizes and reduced defect densities. Our work highlights the transformative potential of agentic learning using visual information in advancing microstructural design, bridging the gap between materials discovery and application readiness.