Cationic perturbation strategy to solve the information drought in material explainable machine learning

In the field of materials research, machine learning (ML) techniques have emerged as indispensable tools. However, the opaqueness in decision making by models can compromise the trustworthiness of results, underscoring the crucial need for model interpretability. Explainable machine learning (XML) strives to augment researchers' comprehension of material properties and performance. Yet, reliance on high-quality datasets and scarcity of prior knowledge pose challenges for XML research, particularly when dealing with smaller datasets. In this study, using spinel as a representative example, we successfully addressed the data challenges in XML through a cationic perturbation strategy. We demonstrate an effective approach for handling information scarcity in small datasets, thus offering a feasible method for material research and broadening the scope of XML applications in materials science. Furthermore, our investigation successfully uncovered potential causal relationships underlying material properties and validated their consistency with physical cognition. These causal relationships can serve as experimental guides, facilitating the design and optimization of new materials. Consequently, this research holds significant scientific merit in advancing XML in the realm of materials science, while providing profound insights into material properties and fostering the development of reliable ML-based materials research.