Machine learning strategies for small sample size in materials science

Abstract

Machine learning (ML) has been widely used to design and develop new materials owing to its low computational cost and powerful predictive capabilities. In recent years, the shortcomings of ML in materials science have gradually emerged, with a primary concern being the scarcity of data. It is challenging to build reliable and accurate ML models using limited data. Moreover, the small sample size problem will remain long-standing in materials science because of the slow accumulation of material data. Therefore, it is important to review and categorize strategies for small-sample learning for the development of ML in materials science. This review systematically sorts the research progress of small-sample learning strategies in materials science, including ensemble learning, unsupervised learning, active learning, and transfer learning. The directions for future research are proposed, including few-shot learning, and virtual sample generation. More importantly, we emphasize the significance of embedding material domain knowledge into ML and elaborate on the basic idea for implementing this strategy.