Abstract Materials informatics is increasingly essential for precise material property prediction, simplifying experimental processes. Traditional methods like density functional theory are sluggish and struggle with complex structure–property relationships. In contrast, artificial intelligence, especially crystal graph convolutional neural networks (CGCNN), excels in predicting material behaviors particularly in crystalline systems revolutionizing materials discovery and design. CGCNN leverages crystal lattice structures for accurate predictions; however, there remains a lack of systematic analysis addressing its foundational principles, performance boundaries, and areas for improvement. This article is novel in offering a comprehensive and critical evaluation of CGCNN, detailing its architecture, predictive strengths, limitations, and integration with emerging technologies such as generative models. It emphasizes benchmarking protocols, best practices, and forward-looking strategies to bridge traditional physics-based methods and modern data-driven approaches. By articulating both the achievements and current gaps in CGCNN-based materials modeling, providing valuable guidance for researchers aiming to harness deep learning for next-generation materials discovery.
Qureshi et al. (Wed,) studied this question.