EXTENDED ABSTRACT: In the era of artificial intelligence, the demand for data in materials genome engineering is continuously increasing. However, in the ffeld of metallic structural materials, the challenge of limited data is particularly acute. To address this issue, we focus on predicting the microstructure and service performance of metallic structural materials by exploring and applying various machine learning (ML) technical strategies. Data augmentation methods can expand limited datasets, enhancing the generalization ability of models. Transfer learning allows us to use abundant source data to pre-train models and transfer this knowledge to smaller datasets, reducing dependence on large amounts of data. Active learning strategies can intelligently select data points with high informational content for labeling, improving learning efffciency. Additionally, explainable ML helps us reveal hidden relationships within material data. to the secretariat as soon as possible. Notably, introducing domain expertise and embedding physical principles into the ML modeling and optimization processes, can achieve an efffcient combination of data-driven and knowledge-driven approaches. This integration not only effectively overcomes the challenges of limited data in materials science, but also helps us deeply understand and predict the behavior of metallic structural materials, providing a more accurate and efffcient new pathway for the novel alloys design.

Keywords: Domain Knowledge Embedded; Physical-informed; Alloys Design

REFERENCES:

[1] Xiong J, Shi S Q, Zhang T Y. Journal of Materials Science & Technology (2021), 87: 133-142.

[2] Xiong J, Zhang T Y. Journal of Materials Science & Technology (2022), 121: 99-104.

[3] Ma J, Cao B, Dong S, et al. npj Computational Materials (2024), 10(1): 59.

[4] Yu Y Y, Xiong J, Wu X, et al. Advanced Science, (2024) 2403548.