<p>Continuous casting machine is a crucial equipment in steelmaking enterprise. Its key continuous casting roller working in cycling temperature and alternating stresses, frequently appeared in wear, crack, bending and so on, directly affecting the quality and efficiency of slab production. Temperature field and thermal stress is viewed as one reason of continuous casting roller failure. It is very difficult to directly capture the surface and internal temperature distribution and stresses under harsh conditions. Therefore, an approach for temperature distribution and stress prediction of continuous casting roller based on finite element modeling (FEM) and machine learning is firstly proposed. The 3D twin model for continuous casting roller is used for constructing corresponding finite element modeling by ANSYS, where the physics-based simulation of temperature and thermal stress distribution is analyzed and obtained. On the basis of simulation-data, temperature and thermal stress prediction model based on support vector regression (SVR) and K-fold cross verification is built for predicting accurately temperature and thermal stress in real-time, thereby effectively reducing the long-simulation time and enhancing the accuracy of thermal stress prediction. Experiment verification shows that proposed approach has higher generalization and real-time performance temperature fields and thermal stresses prediction for continuous casting roller, which can provide digital twin foundation for real-time thermal-state diagnosis and adaptive evolution prediction.</p>