ET-AL: Entropy-Targeted Active Learning for Bias Mitigation in Materials Data

Growing materials data and data-centric informatics tools drastically promote the discovery and design of materials. While data-driven models, such as machine learning, have drawn much attention and observed significant progress, the quality of data resources is equally important but less studied. In this work, we focus on bias mitigation, an important aspect of materials data quality. Quantifying the diversity of stability in different crystal systems, we propose a metric for measuring structure-stability bias in materials data. To mitigate the bias, we develop an entropy-target active learning (ET-AL) framework, guiding the acquisition of new data so that diversities of underrepresented crystal systems are improved, thus mitigating the bias. With experiments on materials datasets, we demonstrate the capability of ET-AL and the improvement in machine learning models through bias mitigation. The approach is applicable to data-centric informatics in other scientific domains.