Transfer Learning‐Based Regression With Interpretable Feature Attribution Predicts Perovskite Morphology for High‐Performance LEDs

ABSTRACT

Perovskite thin‐films have emerged as a pivotal material platform for next‐generation light‐emitting diodes. A deep understanding of how microscopic morphology relates to device performance is essential for advancing perovskite optoelectronics. Although deep learning has achieved classification‐based morphology recognition, it remains challenging to establish a regression relationship between morphology and performance. To overcome this, we develop a domain‐specific pre‐training and transfer learning framework that incorporates an intermediate microstructural database (MicroNet) to bridge the domain gap between natural scene images (ImageNet) and specialized perovskite morphology images. This strategy enables effective learning across domains, benefiting downstream tasks on small‐scale datasets and yielding an accurate morphology‐performance regression model with an R ² of above 0.9. Moreover, we develop a Loss‐guided Class Activation Mapping (CAM) tailored for regression tasks. Unlike conventional CAM, which is primarily designed for classification and visualizes prediction values, our Loss‐guided CAM visualizes the prediction deviation of a regression model by correlating the discrepancy between predicted and true values. Based on this CAM, we conduct a block‐wise analysis and identify microstructural regions with uniform grain size and low porosity as key features, which aligns with domain knowledge and supports the interpretability of our method.