Machine Learning-Guided Discovery of High-Performance Perovskite Solar Cells via Cluster Analysis and Experimental Validation

The optimization of perovskite solar cells (PSCs) is challenged by high-dimensional composition-property relationships in mixed-cation/halide systems. While machine learning (ML) can predict performance, autonomously extracting and validating precise design rules remains difficult. Here, we propose an integrated unsupervised-supervised machine learning framework capable of extraction of microstructural features from scanning electron microscopy (SEM) images to accelerate the discovery of high-efficiency PSCs. The pipeline utilizes Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to design compositional data with morphological descriptors. The predictive model achieves a high predictive accuracy of 97% in distinguishing performance clusters. Guided by this platform, we screened distinct compositional spaces and identified a high-performance "Cluster 0" characterized by FA-dominant compositions (>85%) with low bromine content (<3%). From the 800 AI-generated candidates, we successfully synthesized the top-ranked composition, FA0.94Cs0.03MA0.03Pb(I0.96Br0.04)₃, which exhibited a champion efficiency of 22.06%, aligning closely with the predicted performance. Notably, the proposed method reduces the experimental search space by 3 orders of magnitude. Feature importance analysis confirmed that composition (formamidinium, methylammonium, cesium) is the primary performance driver, experimentally validating our clustering method. This work provides a generalizable tool and a significant roadmap for the data-driven design of complex functional materials, effectively bridging the gap between computational discovery and experimental success.