Integration of Microstructural Image Data into Machine Learning Models for Advancing High-Performance Perovskite Solar Cell Design

Abstract

Perovskite microstructure is one of the key factors limiting the effectiveness of current machine learning (ML) approaches for designing perovskite solar cells (PSCs) with high power conversion efficiency (PCE). This work develops a multimodal convolutional neural network to extract microstructural features from scanning electron microscopy (SEM) images of perovskite thin films. The model dynamically adjusts the weights of different modal information, including material composition, processing techniques, and microstructure, to enhance predictive accuracy. The model achieves an impressive coefficient of determination (R²) of 0.79 on the 1,583 SEM images data set. By introducing six SEM image features to describe the grain size of PSCs, we found that a grain boundary length density (GBLD) below 5.96 and an equivalent circular diameter (ECD) above 0.83 significantly enhance the PCE. Additional experiments confirmed the effectiveness of the results, and by improving these parameters to alter the crystallization, the PCE was increased to 24.61%, and the consistency of the results demonstrated the effectiveness and rationality of the multimodal model.