Metal AM process-structure-property relational linkages using Gaussian process surrogates

Abstract

In metal additive manufacturing (AM), with sufficient understanding of process-structure–property (PSP) relational linkages, the control of build parameters can produce parts with previously unattainable properties. Establishing these PSP linkages involves varying process parameters until a desired microstructure, and corresponding properties, are achieved, either experimentally or computationally. However, both methods can have a high acquisition cost and be difficult to sample repeatedly. This work describes a Gaussian process (GP)-based workflow that is capable of predicting melt pool characteristics, microstructure features, and mechanical properties of previously unseen process parameter combinations. The workflow implements multi-fidelity, multi-output, and functional GPs, trained on a limited set of experiments and simulations in order to make melt pool, microstructure, and mechanical property predictions. The established linkage results in approximately 95% accuracy in predicting mechanical properties for previously unseen set of process parameters propagated through the whole framework. The use of GPs in the workflow limits the number of experiments/simulations needed, yields a nearly negligible cost for acquisition of new predictions, and allows for a Bayesian treatment of the PSP linkages that was not previously feasible.