Explainable machine learning and feature engineering applied to nanoindentation data

Abstract

The work aims to challenge the hegemony in the literature of clustering nanoindentation data solely relying on elastic modulus and hardness as features, thereby discarding information provided by the full load–displacement curve. Features based on dimensional analysis initially aimed to solve the inverse nanoindentation problem were adopted to describe the load–displacement curves. More than 3000 indents in high-speed steels were labelled via imaging after indenting. The resulting dataset was used to train and benchmark supervised (classification) and unsupervised (clustering) machine learning models, showing that feature engineering was more impactful than model selection and hyperparameter tuning, increasing the prediction quality in all studied models. The best classifier’s predictions were explained via a game theory-based approach, allowing insights into the model’s decision-making process and connecting the fields of materials property clustering and materials mechanics.