Evaluating Generalized Feature Importance via Performance Assessment of Machine Learning Models for Predicting Elastic Properties of Materials

Identifying key descriptors and understanding important features across different classes of materials are crucial for machine learning (ML) tools to both predict material properties and reveal the physics underlying any process of interest. Traditionally, the predictive modeling of elastic properties of materials is limited to only a few classes of materials and a small set of ML tools despite the broad applications of these materials. Users now have a broad choice of ML models ranging from simple regression models to graph neural networks (GNN) for predicting structure-property relationships. While in recent years, GNNs have outshined traditional ML models in terms of predictability, their intensive data requirement and lack of interpretability may limit practical applicability. Here, we develop a domain-segmented feature space using a diverse set of material attributes and perform a predictive analysis using state-of-the-art ML tools using elastic modulus prediction as a representative example. By deducing a model-independent overall ranking based on feature importance learned by each model, the knowledge is then transferred to GNNs. Our findings indicate a threshold limit on the predictability of traditional ML models, but our approach of transferring task-specific feature importance knowledge to the GNNs can enhance their performance by reducing their data requirement while retaining considerable accuracy.