Abstract: Against the backdrop of materials genetic engineering, data-driven machine learning (ML) techniques, as a powerful new tool, have garnered widespread attention in the field of research on materials' thermal expansion properties. ML can bypass complex experimental processes and theoretical calculations; by establishing correlations between descriptors and thermal expansion properties, it enables rapid prediction of materials' thermal expansion properties at relatively low cost, effectively compensating for the shortcomings of traditional experimental trial-and-error methods and density functional theory (DFT)-based approaches, such as high time costs and low efficiency. This paper outlines the basic processes and methods of machine learning, and focuses on elaborating its application progress in the research on the thermal expansion properties of materials. The field of traditional machine learning has achieved a gradual deepening, starting from the single-target prediction of the coefficient of thermal expansion, moving to the introduction of an analytical mechanism for feature importance analysis and the combination of feature selection to optimize models, and then advancing to the multi-target prediction associated with multiple properties. In the aspect of machine learning-based interatomic potentials, studies drive molecular dynamics simulations by constructing atomic-scale potential energy functions, thereby revealing the microscopic mechanism of thermal expansion behavior.These applications have accelerated the design and screening of negative thermal expansion materials and deepened the understanding of relevant mechanisms.Finally, the paper analyzes the urgent issues to be resolved in the application of ML to the research on materials' thermal expansion properties, and proposes future research directions and development trends accordingly.

Key words: Machine learning, Thermal expansion prediction, Data-driven, Negative thermal expansion materials