Additive manufacturing, inherent with complex process-structure-property relationships at multiple scales, faces challenges in rapid selection and optimization of manufacturing parameters for desired part quality. Data-driven modeling is a more efficient alternative to experiments and high-fidelity physics-based models [1]. To ensure the reliability and interpretability, we use the simulation results of well-validated high-fidelity models as the training data, and enforce physical laws in the data-driven modelling. We start from data-driven rapid prediction of meso-scale temperature profile, to further implement into micro-scale grain growth simulation and meso-scale thermal stress simulation [2]. The accuracy remains very similar, while the computational speed is significantly increased. Furthermore, the data-driven models serve as strong foundation for data-driven monitoring and diagnostics of additive manufacturing process, which is crucial for intelligent additive manufacturing.