CoSMoR: Decoding decision-making process along continuous composition pathways in machine learning models trained for material properties

Abstract

A key challenge in materials informatics is to decode the decision-making process of machine learning (ML) models that have been trained to predict material properties. The existing methods usually rank alloy features based on importance metrics and do not provide material-specific fundamental insights. Here, we present the Compositional Stimulus and Model Response (CoSMoR) framework that can be applied to any composition-based ML model (irrespective of the algorithm used) to calculate the exact contribution of each feature towards the manifestation of target material property along a continuous compositional pathway. CoSMoR utilizes the local partial dependencies of target property with respect to each feature and combines it with feature variations associated with discretized compositional variations to measure exact feature contributions. We showcase the importance of CoSMoR through implementation on phase-selection problem in multiprincipal element alloys (MPEAs), wherein it leads to physical insights into phase transitions. A detailed overview of the framework, along with the codes and step-by-step implementation of the algorithm, has been provided to enable extension to new or preexistent models.