EXTENDED ABSTRACT: The machine learning is accelerating the development of high-performance steel materials by bypassing complex physical mechanisms, thus avoiding issues like parameter sensitivity and error accumulation. Traditional methods, such as automated laboratories, have emerged but suffer from poor interpretability and data dependency. This study proposes a steel material gene engineering method that integrates expert physical knowledge with data mining, leveraging the strong data mining capabilities of machine learning and the clear mechanisms of physical models. This method introductions physical metallurgy information or real microstructural data to accurately predict and design steel material properties, demonstrating high tolerance for small sample sizes. Firstly, thermodynamic data is used to model the relationship between solid-state phase transformations and strength. Under the guidance of thermodynamic data, deep learning mechanisms are employed to develop precise predictive models for friction work and Ms temperature in multicomponent systems. For more complex creep/fatigue issues, source models are established to learn the relationship between composition, process, and creep/fatigue either directly or through conventional mechanical properties. These learned mechanisms are transferred to new models, enabling accurate predictions and design of the fatigue/creep. However, these parameters often fail to fully capture real interactions in more complex problems. Therefore, this study further introduces microstructural information to guide AI learning. Multiple rapid quantiffcation methods for microstructural images are developed. Additionally, multimodal information from materials science is used to guide AI strategies, achieving comprehensive integration of materials analysis methods and informatics models, thereby enhancing the accuracy of performance prediction models.

Keywords: Physical metallurgical information; Microstructure; Steel materials; Machine learning;