Perovskite solar cells (PSCs) are at the forefront of next-generation photovoltaics, combining high power conversion efficiency with low-cost and scalable fabrication. Their unique optoelectronic properties and potential for flexible, lightweight applications make them strong candidates for future solar energy systems. However, key challenges, particularly related to long-term stability, scalable manufacturing, and consistent film quality, must be addressed to enable commercial deployment [1,2]. To tackle these issues, we present an automated platform that integrates machine learning (ML) workflows with high-resolution image analysis for real-time process optimization. As illustrated in Figure 1, this platform combines computational tools and data-driven methods to study and impove the crystallization dynamics and morphological quality of PSC films.

Our approach leverages microscopic imaging acquired in situ or post-deposition, which is processed using state-of-the-art segmentation frameworks, including the Segment Anything Model (SAM) and Detectron2. These tools enable the precise identification of morphological features such as crystal domains, grain boundaries, and nuclei formation sites. To further classify and characterize these regions, we employ a ResNet152 convolutional neural network (CNN), which supports detailed recognition of grain size, shape, and orientation.
From the segmented images, we extract a set of quantitative descriptors to assess film homogeneity and microstructural quality. These include crystal size distributions, aspect ratios, and novel information-theoretic metrics such as Shannon entropy and Computable Information Density (CID). These metrics enable a deeper understanding of the crystallization process, revealing correlations between microstructural patterns and device performance. [3].
Applied across multiple fabrication batches, our pipeline captures subtle morphological variations and identifies key trends linked to processing conditions. In particular, we show that entropy and CID serve as sensitive, compact indicators of film homogeneity, complementing physical measurements and aiding in the detection of suboptimal growth regimes.

This methodology provides a scalable and automated framework for image-based quality assessment in PSC fabrication. While real-time control is not yet implemented, the demonstrated analytical capacity lays the groundwork for future integration into closed-loop synthesis platforms, ultimately supporting more consistent and efficient manufacturing.