Abstract
The development of vapor deposition technology will accelerate the process of perovskite solar cells (PSCs) moving from laboratory scale to industrialization. However, the multi-dimensional and complex parameter space will inevitably increase the cost of trial and error, especially for high-energy-consuming and long-cycle vapor deposition technologies. This study employed an integrated-feature dataset encompassing macro- and micro-features to enhance the accuracy, robustness, and interpretability of an ETree machine learning (ML) model for power conversion efficiency (PCE) prediction, achieving an impressive coefficient of determination value of 0.9464 and root mean square error value of 1.27%. Through SHAP analysis, Monte Carlo simulations, and parameter space exploration, an optimal FTO/SnO(2)/Cs(0.04)FA(0.96)PbI(3)/Spiro-OMeTAD/Au device architecture and vapor deposition parameters are reverse-engineered, yielding the highest predicted PCE of 26.21%. Furthermore, the PCEs are enhanced by implementing the universal ML-derived optimization strategies across six distinct and independent vapor deposition processing, which truly realizing the objective of ML-guided experiments based on various preparation conditions. This machine learning model is believed to shorten the research and development cycle for breaking through the performance bottleneck of high-efficiency devices fabricated by vapor deposition technology, providing a potential approach for its commercial application.