Session: DAC-07-01-Design for Additive Manufacturing

Paper Number: 89091

89091 - Deep Ensembles for Modeling Uncertain Phase Constraints In Compositionally Graded Alloy Design 

Compositionally graded alloys are a specific class of multi-material functionally graded materials (FGMs) that use spatial variations in alloy composition to meet competing performance requirements in different regions of a single part. Previously, the authors presented a computational design methodology that optimizes for user-input performance objectives while solving common issues with designing compositionally graded alloys, including deleterious phase formation and undesirable property profiles. While the methodology has been used to successfully design gradient alloys from conception to manufacturing, the k-nearest neighbor (KNN) CALPHAD surrogate models used for constraint modeling have no uncertainty metric, which can result in uncertain model(s) over-constraining the design space to the point where narrow passageways, and feasible design regions in general, are either constricted or eliminated entirely. These regions can be especially critical in building an optimal gradient path, and, thus, these true free regions being mistakenly constrained can result in anything from sub-optimal designs at best to complete design opportunities lost at worst. Since the models and design state spaces are too complex for a manual analysis of the modeling constraint boundaries, there is a need for a modeling technique that performs as well or better than the current KNN approach while also providing information about the prediction uncertainty. In this work, the previous methodology is improved by proposing the use of a deep ensemble for CALPHAD surrogate phase constraint modeling. The performance of a tuned deep neural network (NN) and tuned deep ensemble is first analyzed by testing them against a benchmark NN and the currently-used KNN algorithm in a material test problem. Then, based on the admirable performance of the tuned deep ensemble in the initial study, the same model is applied to a complete materials data set, and the predicted class probability threshold is manipulated to create acceptable phase regions with both higher and lower degrees of uncertainty, to illustrate the varied design preferences of a given user. The study results show that these restricted and relaxed phase regions can eliminate and create narrow passageways, respectively, indicating the usefulness of this uncertainty metric in tuning the design space. Lastly, there is a discussion regarding further work in this area, such as testing the tuned ensemble on a wider variety of material systems and leveraging the model uncertainty to implement an adaptive sampling method to further improve the computational FGM design methodology.

Presenting Author: Marshall Allen Texas A&M University

Presenting Author Biography: Marshall Allen is a Ph.D. student in the Department of Mechanical Engineering at Texas A&M University. He is co-advised by Dr. Richard Malak and Dr. Raymundo Arroyave. Marshall’s research focus is refining the use of robotic path planning algorithms for the design of Functionally Graded Materials (FGMs).

Authors:

Marshall Allen Texas A&M University
Raymundo Arroyave Texas A&M University
Richard Malak Texas A&M University