Machine learning strategies for small sample size in materials science  

材料科学中小样本量的机器学习策略

作　　者：Qiuling Tao JinXin Yu Xiangyu Mu Xue Jia Rongpei Shi Zhifu Yao Cuiping Wang Haijun Zhang Xingjun Liu 陶秋伶;于金鑫;穆翔宇;贾雪;施荣佩;姚志富;王翠萍;张海军;刘兴军(School of Materials Science and Engineering,and Institute of Materials Genome&Big Data,Harbin Institute of Technology,Shenzhen,Shenzhen 518055,China;School of Computer Science&Technology,Harbin Institute of Technology,Shenzhen,Shenzhen 518055,China;Advanced Institute for Materials Research(WPI-AIMR),Tohoku University,Sendai 980-8577,Japan;Department of Materials Science and Engineering,College of Materials,Xiamen University,Xiamen 361005,China;State Key Laboratory of Advanced Welding and Joining,Harbin Institute of Technology,Shenzhen,Shenzhen 518055,China)

机构地区：[1]School of Materials Science and Engineering,and Institute of Materials Genome&Big Data,Harbin Institute of Technology,Shenzhen,Shenzhen 518055,China [2]School of Computer Science&Technology,Harbin Institute of Technology,Shenzhen,Shenzhen 518055,China [3]Advanced Institute for Materials Research(WPI-AIMR),Tohoku University,Sendai 980-8577,Japan [4]Department of Materials Science and Engineering,College of Materials,Xiamen University,Xiamen 361005,China [5]State Key Laboratory of Advanced Welding and Joining,Harbin Institute of Technology,Shenzhen,Shenzhen 518055,China

出　　处：《Science China Materials》2025年第2期387-405,共19页中国科学(材料科学)(英文版)

基　　金：supported by the National Natural Science Foundation of China (52371007 and 52301042);the National Key R&D Program of China (2020YFB0704503);the Guangdong Basic and Applied Basic Research Foundation (2021B1515120071);the Key-Area Research and Development Program of Guangdong Province (2023B0909050001)。

摘　　要：Machine learning (ML) has been widely used todesign and develop new materials owing to its low computational cost and powerful predictive capabilities. In recentyears, the shortcomings of ML in materials science have gradually emerged, with a primary concern being the scarcity ofdata. It is challenging to build reliable and accurate ML modelsusing limited data. Moreover, the small sample size problemwill remain long-standing in materials science because of theslow accumulation of material data. Therefore, it is importantto review and categorize strategies for small-sample learningfor the development of ML in materials science. This reviewsystematically sorts the research progress of small-samplelearning strategies in materials science, including ensemblelearning, unsupervised learning, active learning, and transferlearning. 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 thebasic idea for implementing this strategy.机器学习(ML)因其较低的计算成本和强大的预测能力,已广泛应用于新材料的设计与开发.近年来,机器学习在材料科学中的不足逐渐显现,其中最主要的问题是数据稀缺.在有限数据的情况下,构建可靠且准确的机器学习模型是一项挑战.此外,由于材料数据积累的速度较慢,小样本量问题将长期存在于材料科学中.因此,回顾并分析针对小样本学习的策略,对促进机器学习在材料科学中的发展具有重要意义.本文系统地梳理了材料科学中小样本学习策略的研究进展,包括集成学习、无监督学习、主动学习和迁移学习.并提出了未来研究的方向,如少样本学习和虚拟样本生成.更为重要的是,我们强调了将材料领域知识嵌入机器学习的重要性,并阐述了实现该策略的基本思路.

关 键 词：material design machine learning small sample size few-shot learning material domain knowledge 

分 类 号：H31[语言文字—英语]