Enhancing interpretability in the exploration of high-energy conversion efficiency in CsSnBr3−xIx configurations using crystal graph convolutional neural networks and adversarial example methods

Crystal graph convolutional neural networks (CGCNNs) have revolutionized materials research by eliminating the need for manual feature engineering. However, their lack of interpretability and sensitivity to structural distortions hinders their application in substitution engineering. Therefore, we propose an innovative adversarial example method for guiding feature construction and enhancing the interpretability of CGCNNs. In this study, our focus lies on identifying CsSnBr3−xIx configurations with high-energy conversion efficiency. Initially, we train a CGCNN classifier as a benchmark. Subsequently, we perturb input data to generate a low-performance classifier and identify adversarial examples based on incorrect predictions. Upon comparing these examples with normal examples, we observe substantial structural distortions in adversarial cases, serving as inspiration for the creation of disorder-related features. Consequently, an interpretable model is developed, which surpasses CGCNNs with atomic position perturbations and a gradient-boosting classifier using general features. Notably, the previously overlooked feature “number of unequal atoms” plays an important role in offering crucial insights. Further analysis reveals that configurations with pronounced disorder can exhibit increased power density, thereby enhancing the energy conversion efficiency. Our work not only elucidates the impact of atom substitution on energy conversion efficiency but also provides a roadmap for constructing interpretable machine learning models.