CrabNet for explainable deep learning in materials science: bridging the gap between academia and industry

Despite recent breakthroughs in deep learning for materials informatics, there exists a disparity between their popularity in academic research and their limited adoption in the industry. A significant contributor to this “interpretability-adoption gap” is the prevalence of black-box models and the lack of built-in methods for model interpretation. While established methods for evaluating model performance exist, an intuitive understanding of the modeling and decision-making processes in models is nonetheless desired in many cases. In this work, we demonstrate several ways of incorporating model interpretability to the structure-agnostic Compositionally Restricted Attention-Based network, CrabNet. We show that CrabNet learns meaningful, material property-specific element representations based solely on the data with no additional supervision. These element representations can then be used to explore element identity, similarity, behavior, and interactions within different chemical environments. Chemical compounds can also be uniquely represented and examined to reveal clear structures and trends within the chemical space. Additionally, visualizations of the attention mechanism can be used in conjunction to further understand the modeling process, identify potential modeling or dataset errors, and hint at further chemical insights leading to a better understanding of the phenomena governing material properties. We feel confident that the interpretability methods introduced in this work for CrabNet will be of keen interest to materials informatics researchers as well as industrial practitioners alike.