Microstructure Generation via Generative Adversarial Network for Heterogeneous, Topologically Complex 3D Materials
- PDF / 2,500,060 Bytes
- 13 Pages / 593.972 x 792 pts Page_size
- 79 Downloads / 178 Views
https://doi.org/10.1007/s11837-020-04484-y Ó 2020 The Minerals, Metals & Materials Society
AUGMENTING PHYSICS-BASED MODELS IN ICME WITH MACHINE LEARNING AND UNCERTAINTY QUANTIFICATION
Microstructure Generation via Generative Adversarial Network for Heterogeneous, Topologically Complex 3D Materials TIM HSU,1,2,6 WILLIAM K. EPTING,1,3 HOKON KIM,1,2 HARRY W. ABERNATHY,4,5 GREGORY A. HACKETT,4 ANTHONY D. ROLLETT,1,2 PAUL A. SALVADOR,1,2 and ELIZABETH A. HOLM 1,2,7 1.—US DOE National Energy Technology Laboratory, Pittsburgh, PA 15236, USA. 2.—Materials Science and Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA. 3.—Leidos Research Support Team, Pittsburgh, PA 15236, USA. 4.—US DOE National Energy Technology Laboratory, Morgantown, WV 26505, USA. 5.—Leidos Research Support Team, Morgantown, WV 26505, USA. 6.—Lawrence Livermore National Laboratory, Livermore, CA 94550, USA. 7.—e-mail: [email protected]
Using a large-scale, experimentally captured 3D microstructure data set, we implement the generative adversarial network (GAN) framework to learn and generate 3D microstructures of solid oxide fuel cell electrodes. The generated microstructures are visually, statistically, and topologically realistic, with distributions of microstructural parameters, including volume fraction, particle size, surface area, tortuosity, and triple-phase boundary density, being highly similar to those of the original microstructure. These results are compared and contrasted with those from an established, grain-based generation algorithm (DREAM.3D). Importantly, simulations of electrochemical performance, using a locally resolved finite element model, demonstrate that the GAN-generated microstructures closely match the performance distribution of the original, while DREAM.3D leads to significant differences. The ability of the generative machine learning model to recreate microstructures with high fidelity suggests that the essence of complex microstructures may be captured and represented in a compact and manipulatable form.
INTRODUCTION As a component of integrated computational materials engineering (ICME), methods to generate realistic simulation volumes are essential for modeling and simulating materials with complex microstructures.1 While obtaining microstructures experimentally guarantees realistic simulation volumes, the cost or difficulty of microstructural characterization often limits the size or number of microstructures that can be sampled, particularly in 3D. Thus, to support computational design and performance surveys, a common goal is to synthesize statistically representative sets of microstructural realizations.2 A number of successful approaches have been developed, such as those based on n-point correlation ellipsoid packing,5,6 physical functions,3,4 (Received June 24, 2020; accepted October 28, 2020)
descriptors,7 Gaussian random field,8 and Markov random field.9 Bostanaband et al.10 provide an extensive review of existing microstructure reconstruction techniques. However, despite their succes
Data Loading...
