Crystallization of perovskites is the key process in thin film perovskite solar cells (PSCs). The ability to connect film formation dynamics with device performance remains a central challenge in the development of high-performance PSCs. Here, we develop a deep spatiotemporal attention network that outputs a current density-voltage (J-V) curve for each in-situ photoluminescence (PL) patch. Training follows a weak-supervision paradigm: only the global J-V is observed, and learning is driven by aligning the aggregation of local predictions with this global target. Visual Explanation analysis via gradient-based activation map (GAM) indicates that the model focuses on physically meaningful stages of film evolution—capturing early nucleation dynamics and later crystallization—while emphasizing longer-wavelength PL signals during late stages and shorter-wavelength cues during nucleation. Beyond global prediction, the local J-V outputs enable spatial diagnostics: using Spearman’s rank correlation, we find a statistically significant negative monotonic association between the dispersion of local power conversion efficiency (PCE) and overall PCE, which strengthens on well-estimated samples. Together, these results suggest that spatially uniform local performance is a strong indicator of high-performance device and highlight PL-video-driven deep learning as a scalable data-driven workflow for both performance prediction and process optimization in industry. To demonstrate industrial applicability of solar cells, this AI model integrates into scalable production lines like blade-coating. By rapidly evaluating uniformity and predicting performance, it enables a real-time feedback loop for immediate process adjustments, ensuring consistent, high-quality manufacturing. • A model based on Transformer architecture and weak supervision is developed to enable the rapid and precise prediction of the local performance of a perovskite film. • The results demonstrate a marked improvement in predictive capability: The prediction accuracy improves from 0.472 to 0.504. The prediction error is substantially reduced, with the Mean Absolute Error decreasing from 1.488 to 1.358. • A clear negative correlation emerges: devices with lower global Power Conversion Efficiency (PCE) tend to exhibit higher spatial variability in the predicted local efficiencies, whereas devices with higher global PCEs display more uniform local predictions.
Lin et al. (Sun,) studied this question.