Development and application of Few-shot learning methods in materials science under data scarcity

Abstract

Machine learning, as a significant branch of artificial intelligence, has provided effective guidance for material design by establishing virtual mappings between data and desired features, thereby reducing the cycle of material discovery and synthesis. However, the application of machine learning in materials science is hindered by data scarcity. Few-shot learning methods, an effective approach for improving the performance of machine learning models under data scarcity, have achieved significant development in the field of materials science. In this review, the recent advancements in few-shot learning methods in materials science are discussed, and the application workflow of machine learning algorithms is elucidated. Methods for dataset expansion are discussed from the perspective of data acquisition, including databases, natural language processing, and high-throughput experiments, while collating commonly used materials science databases in the process. The application of algorithms, such as transfer learning and data augmentation in materials science, was analyzed in few-shot environments in materials science. Finally, the challenges faced by the application of machine learning in materials science are summarized, and the related future prospects are outlined.