Deep-Learning-Based In-Situ Photoluminescence and Reflection Monitoring for Property Prediction of Perovskite Solar Cells

The reproducible and uniform fabrication of high-quality, large-area perovskite thin films remains a significant challenge for the commercialization of perovskite solar cells (PSCs). This work demonstrates how deep learning (DL) enhances the capabilities of in-situ photoluminescence (PL) and diffuse reflection imaging to monitor and optimize the perovskite thin-film formation process. By analyzing high-dimensional, time-resolved data, acquired in-situ during vacuum quenching, with DL, we advance the ability to predict and control the fabrication of PSCs. Our approach focuses on leveraging DL to enhance the predictive power of PL and reflection imaging. First, neural networks are trained to monitor material composition, identifying subtle variations in precursor molarity and molar ratios, which is difficult to achieve with human analysis. This ensures material consistency and enables early detection of potential fabrication defects. Second, various machine learning models are used to predict both thin film thickness and device performance directly from in-situ data. This enables early in-line assessment of thin film quality, providing insight into device-level properties prior to solar cell completion. Finally, the predictive approach is extended beyond passive monitoring by integrating predictions into a DL-driven process control system. This system forecasts trends in PL monitoring data and predicts corresponding device performance, providing actionable recommendations for real-time process optimization. By augmenting in-situ PL and reflection imaging with DL, this work paves the way for enhanced monitoring of the scalable production of PSCs.