Abstract Scope |
The lack of complete process-structure-property (P-S-P) relationships for metal additive manufacturing is still the bottleneck in producing defect-free, structurally sound, and reliable parts. To alleviate the curse of dimensionality in constructing P-S-P relationships, a physics- constrained machine learning approach is proposed to construct surrogates in a high-dimensional parameter space with reduced amount of training data. In this work, the physics-constrained neural network with the minimax architecture (PCNN-MM) is developed to predict the dendritic growth of Ti-6Al-4V alloy during the rapid solidification process. The training of the PCNN-MM is to solve a minimax problem by searching the saddle points of the objective function with a Dual-Dimer saddle point search algorithm. The results show that the PCNN-MM can provide fast online predictions of phase field, composition field, and temperature field after offline training. The PCNN-MM has the potential of accelerating materials design and process design where simulation and experimental data are sparse. |