Subgroup Discovery Similarity Score (SDSS): A Significant Criterion for the Integration of Statistical Knowledge into Machine Learning in Materials Science

Abstract

In materials science research, knowledge and machine learning (ML) have a mutually reinforcing relationship. In efforts to improve the ability of learning material datasets, researchers obtain statistical knowledge from ML models and integrate it into subsequent ML models in different ways. However, determining the most suitable method for integrating statistical knowledge into the next stage remains challenging. This limits the precise application of knowledge-driven approaches. In this work, the Subgroup Discovery Similarity Score (SDSS) is proposed as a key criterion for integrating statistical knowledge into ML models. Statistical knowledge is extracted from material datasets by subgroup discovery. In the solid solution strengthening ($$) dataset, a divide-and-conquer strategy achieves a correlation coefficient of 0.96 and a MAPE of 18.44%, and reveals distinct strengthening mechanisms for the face-centered cubic (FCC) and body-centered cubic (BCC) phases. In the piezoelectric coefficients ($$) dataset, statistical knowledge is encoded as features and embedded into the ML model for feature enhancement, effectively reducing the prediction error. The results suggest that our framework can extract and integrate statistical knowledge from material datasets into ML models without prior domain knowledge.