Alex Kitt
Luke Mohr
Arushi DhakadPredicting AM-Part Performance
August 23, 2021Comments
The additive manufacturing (AM) build process is exciting and state of the art, and enables manufacture of new classes of parts while changing how we think about supply chains. However, focusing on the build process can undervalue the importance of post-processing on AM-part performance. Post-processing steps such as heat treatment, hot isostatic pressing (HIP) and surface finishing, though not as compelling as building a part one layer at a time, significantly affect AM-part performance. Recognizing the importance of post-processing, we’ll describe a probabilistic model developed by EWI to predict part performance based on both the build process and the post-processing technique.
During the last three years, EWI’s Additive Manufacturing Consortia (AMC) ran a series of programs that quantified aspects of AM surface finishing. The projects were chosen, guided and funded by the more than 50 AMC industry members and supported by various AM surface-finishing-equipment manufacturers. The 2020 program investigated the roughness of as-built parts; the 2018 program quantified the improvement in roughness afforded by different surface-finishing approaches; and the 2019 program measured the effect of roughness on fatigue. Together, these programs provide members with a rich dataset spanning the causal process chain from build to post-processing to performance.
EWI has used this data to train a predictive model for part performance. This model, a generative Bayesian network (Fig. 1), consists of 3D-separable subgraphs trained using Bayesian linear regression. Bayesian networks provide a machine-learning method for forward and backward prediction of causal processes. The model includes the effects of surface angle, orientation of the surface relative to the rake and gas flow, position of the coupon within the build envelope, layer thickness, surface-finishing technique and stress level for fatigue testing. The as-built roughness data come from Ti64 components built using laser powder-bed fusion while the second two datasets result from Ti64 components built using electron beam melting with HIP. We assume transferability given the consistency in materials, but additional data can be used to establish a discrepancy function. In this article we will summarize three usecases for this model.
A what-if scenario (Fig. 2) demonstrates tradeoffs between design and surface-finishing technique. In the as-built condition (Fig. 2, left) the predicted fatigue performance for a 45-deg. downskin (red) is slightly worse than that of a vertical wall (blue)—consistent with rougher downskin surfaces resulting in earlier crack formation. The roughness reduction provided by the REM ISF process (an extreme isotropic surface-finishing technique) is predicted to improve the fatigue performance for both surface angles (Fig. 2, right). However, note a more pronounced performance difference between 45 and 90 deg. (vertical) due to the fatigue model being more sensitive in the lower roughness ranges. We ran this scenario for a 30-micron layer thickness with orientation and position in the build envelope as uniform random variables. Alternative scenarios can be modeled in real time to further optimize performance.
As a second use case of the forward model, we consider the selection of a quality-control plan. Here, we compare two plans designed to ensure a minimum fatigue life of 10,000 cycles at 1-percent maximum strain. In the first strategy, we inspect the as-built samples. This plan can reduce costs by rejecting parts before surface finishing. In the second plan, we inspect the parts in the surface-finished state. Fig. 3 shows two histograms comparing the predicted fatigue life at 1-percent maximum stress for a component finished via the REM ISF process with a 30-micron layer thickness, a 90-deg. surface angle and all other input variables unconstrained. 
As an example, consider a minimum life requirement of 1000 cycles at 1-percent max strain, 100 cycles at 2-percent max strain and 10 cycles at 3-percent max strain. First evaluate the parameters with the biggest influence: surface-finishing technique, layer thickness and surface angle, while leaving the remaining parameters unconstrained. The selection of surface-finishing process proves critical for achieving a 95-percent probability of success (Fig. 4, left). To further improve the probability of success, we consider the position and orientation of the part for the REM ISF surface-finishing process, a 60-micron layer thickness, and surface angle between 75 and 105 deg. We show results with the build area broken into a 3-by-3 grid (Fig. 4, right), with a polar plot of orientation plotted for each grid location. The red line corresponds to a 96.5-percent probability of success, illustrating that careful selection within the full process window can further improve the probability of success.
Video