Interpretable ML for Materials

The ML approaches discussed thus far focused on the developing composition-property modelsModels or to use these modelsModels toward the inverse design of materials. However, understanding the nature of these black-box modelsModels are important so that an informed decision making process can be employed. To this extent, in this chapter, interpretable ML modelsModels are discussed. First, SHAP, a post-hoc modelModels agnostic approach is employed to interpret the composition property modelModels. SHAP provides insights into the features governing a property both in a qualitative and quantitative manner. Further, SHAP also provides the coupling or interaction between the input features for a given property. Finally, the use of support vector machines to interpret the structure–dynamics relationships in materials through a novel machine-learned metric, namely, “softness” is discussed. Altogether, the chapter outlines how interpretable ML can be used to gain insights into the black-box functions for materials response.