SUMMARY

Metal additive manufacturing (AM) is a group of processes by which metal parts are built layer by layer from powder or wire feedstock with high-energy laser or electron beams. The most well-known metal AM processes include selective laser melting, electron beam melting, and direct energy deposition. Metal AM can significantly improve the manufacturability of products with complex geometries and heterogeneous materials. It has the potential to be widely applied in various industries. However, the lack of complete process-structure-property (P-S-P) relationships for metal AM is still the bottleneck to produce defect-free, structurally sound, and reliable AM parts. There are several technical challenges in establishing the P-S-P relationships for process design and optimization. First, there is a lack of fundamental understanding of the rapid solidification process during which microstructures are formed and the properties of solid parts are determined. Second, the curse of dimensionality in the process and structure design space leads to the lack of data to construct reliable P-S-P relationships. Simulation becomes an important tool to enable us to understand the rapid solidification given the limitations of experimental techniques for in-situ measurement. In this research, a mesoscale multiphysics simulation framework, called phase-field and thermal lattice Boltzmann method (PF-TLBM), is developed with simultaneous considerations of heterogeneous nucleation, solute transport, heat transfer, and phase transition. The simulation can reveal the complex dynamics of rapid solidification in the melt pool, such as the effects of latent heat and cooling rate on dendritic morphology and solute distribution. The microstructure evolution in the complex heating and cooling environment in the layer-by-layer AM process is simulated with the PF-TLBM framework. To meet the lack-of-data challenge in constructing P-S-P relationships, a new scheme of multi-fidelity physics-constrained neural network (MF-PCNN) is proposed to improve the efficiency of training in neural networks by reducing the required amount of training data and incorporating physical knowledge as constraints. Neural networks with two levels of fidelities are combined to improve the prediction accuracy. The proposed MF-PCNN is applied to predict phase transition and dendritic growth in AM. The training of MF-PCNN is also formulated as a minimax problem, and a novel training algorithm called Dual-Dimer is developed to search high-order saddle points. A surrogate model of process-structure relationship for AM is constructed based on the PF-TLBM and MF-PCNN. Based on the surrogate model, multi-objective Bayesian optimization is utilized to search the optimal initial temperature and cooling rate to obtain the desired dendritic area and microsegregation level. The proposed PF-TLBM and MF-PCNN provide a systematic and efficient approach to construct P-S-P relationships for AM process design.