Metal Additive Manufacturing Process Design based on Physics Constrained Neural Networks and Multi-Objective Bayesian Optimization

Abstract

The high-dimensionality of the process-structure-property (P-S-P) relationships for metal additive manufacturing (AM) is the major research challenge for process optimization to produce defect-free, structurally sound, and reliable AM parts. The goal of this work is to investigate the feasibility of a hybrid physics-based data-driven process design framework to establish reliable surrogates of process-structure relationships for metal AM process optimization. The process design framework includes a mesoscale multiphysics simulation model to predict microstructure evolution, a physics-constrained neural network to construct the surrogate of the process-structure relationship, and Bayesian optimization for process design. The proposed framework is demonstrated by optimizing the initial temperature and cooling rate for the single dendritic growth of Ti-6Al-4V alloy during rapid solidification in metal AM so that the desired dendritic area and microsegregation level can be achieved. The multi-objective optimization problem is solved with single-objective Bayesian optimization by aggregation. The effects of the aggregation weights on the optimization results are investigated with sensitivity analysis. The results show that the weights of individual objective functions determined from design preferences can help the designer to produce desired microstructure. The Pareto front and processing map for dendritic growth are constructed with the multi-objective Bayesian optimization method. The proposed process design framework is generic and can be potentially applied in materials design and digital twins of metal AM.