Automated and Scalable SEM Image Analysis of Perovskite Solar Cell Materials via a Deep Segmentation Framework

TL;DR

This work presents PerovSegNet, a high-precision SEM image segmentation framework built on YOLOv8-seg for perovskite and lead iodide films. It introduces two novel modules, ASDCB for multi-scale feature fusion and boundary localization, and SAD for preserving fine textures and large-scale morphology, achieving state-of-the-art segmentation on the PerovData dataset with efficient compute. Beyond segmentation, it outputs grain-level metrics (areas and counts) that correlate with crystallization quality and device performance, enabling data-driven process optimization. The approach supports scalable, real-time monitoring and can accelerate high-throughput screening and inline quality control in perovskite manufacturing, contributing to improved efficiency and stability of PSC devices.

Abstract

Scanning Electron Microscopy (SEM) is indispensable for characterizing the microstructure of thin films during perovskite solar cell fabrication. Accurate identification and quantification of lead iodide and perovskite phases are critical because residual lead iodide strongly influences crystallization pathways and defect formation, while the morphology of perovskite grains governs carrier transport and device stability. Yet current SEM image analysis is still largely manual, limiting throughput and consistency. Here, we present an automated deep learning-based framework for SEM image segmentation that enables precise and efficient identification of lead iodide, perovskite and defect domains across diverse morphologies. Built upon an improved YOLOv8x architecture, our model named PerovSegNet incorporates two novel modules: (i) Adaptive Shuffle Dilated Convolution Block, which enhances multi-scale and fine-grained feature extraction through group convolutions and channel mixing; and (ii) Separable Adaptive Downsampling module, which jointly preserves fine-scale textures and large-scale structures for more robust boundary recognition. Trained on an augmented dataset of 10,994 SEM images, PerovSegNet achieves a mean Average Precision of 87.25% with 265.4 Giga Floating Point Operations, outperforming the baseline YOLOv8x-seg by 4.08%, while reducing model size and computational load by 24.43% and 25.22%, respectively. Beyond segmentation, the framework provides quantitative grain-level metrics, such as lead iodide/perovskite area and count, which can serve as reliable indicators of crystallization efficiency and microstructural quality. These capabilities establish PerovSegNet as a scalable tool for real-time process monitoring and data-driven optimization of perovskite thin-film fabrication.The source code is available at:https://github.com/wlyyj/PerovSegNet/tree/master.

Figures (11)

• Figure 1: Overview of the PerovSegNet-based segmentation workflow. The pipeline begins with SEM image acquisition, followed by the evaluation of structural separability across molecular species using UMAP. A ROI is selected and annotated to construct a training dataset, which is then augmented and used to train PerovSegNet. The trained model performs automated segmentation and predicts pixel-wise classifications for perovskite, lead iodide and defect.
• Figure 2: Dataset construction and separability evaluation. a) A representative SEM image from PerovData dataset, captured at 20,000 $\times$ magnification, reveals the granular morphology of perovskite films. b) Annotated segmentation map showing individual regions of perovskite (green), lead iodide (red), and defects (blue). c) UMAP projection of the dataset highlighting molecular separability, where clusters correspond to perovskite (purple), lead iodide (dark green), and defects (light green).
• Figure 3: Statistical analysis of annotation distribution. a) Instance frequency across categories: lead iodide, perovskite, and defect. b) Heatmap showing the size and number of instances across SEM image regions. c) Scatter plot of the position distribution of the center point of instances relative to the entire image. d) Joint distribution of bounding box width and height.
• Figure 4: Overall architecture of PerovSegNet. The network consists of three stages: Backbone for hierarchical feature extraction, Neck for multi-scale feature fusion, and Head for segmentation map generation. PerovSegNet tailored for SEM image segmentation of perovskite and lead iodide, extends the YOLOv8x-seg framework by incorporating two customized modules: ASDCB and SAD, which enhance sensitivity to fine-grained morphology and robustness against complex backgrounds.
• Figure 5: Architecture of the ASDCB module. The module is composed of four parallel branches: Standard Convolution, Dilated Convolution, SE layer, and Depthwise Separable Convolution. Each branch is specifically designed to extract complementary features across different scales. The outputs from these branches are fused through a Channel Shuffle operation, which is followed by a ConvFFN for feature integration and transformation. This design enables ASDCB module to enhance the feature representation for complex SEM images of perovskite and lead iodide materials.