Data-driven prediction of keyhole features in metal additive manufacturing based on physics-based simulation

The defect formation is closely related to molten pool and keyhole features in metal additive manufacturing. Experimentation and physics-based simulation methods to capture the molten pool and keyhole features are expensive and time-consuming. A data-driven method is proposed in this work to efficiently predict the molten pool and keyhole features characterized by a series of fitting curves under given manufacturing parameters, instead of simply predicting the molten pool and keyhole sizes. The database consists of simulation cases with the high-fidelity thermal-fluid flow model. Molten pool melting regime, keyhole stability and keyhole type are recognized with the neural net pattern recognition. With the Gaussian process regression model, the keyhole dimensions are predicted and the keyhole contour is reconstructed. The comparison between predicted data and physics-based simulation results demonstrates the feasibility and accuracy of our data-driven model. Meanwhile, the predicted results can guide the selection of manufacturing parameters in actual production, and are also helpful to the further study of pores in additive manufacturing in academic research.