Source-linked AI summary
Microstructure Generation via Generative Adversarial Network for Heterogeneous, Topologically Complex 3D Materials
Tim Hsu, William K. Epting, Hokon Kim, Harry W. Abernathy, Gregory A. Hackett, Anthony D. Rollett, Paul A. Salvador, Elizabeth A. Holm
TL;DR
Experimental 3D microstructure characterization is costly, and many reconstruction methods assume restricted structural forms. This paper trains a GAN on experimentally captured SOFC anode volumes and compares its generated structures with experimental data and DREAM.3D. The GAN structures closely match experimental electrochemical performance distributions, while DREAM.3D differs significantly.
Problem
Experimental characterization can limit the size or number of 3D microstructures available, while many reconstruction methods assume restricted structures.
Source: §1 · pp. 1–2Method
A deep-convolutional GAN learns from experimentally captured 3D SOFC anode microstructures and generates synthetic volumes.
Source: §1 · p. 2, §2.1 · p. 3, §2.3 · p. 4Results
GAN-generated microstructures closely matched the experimental electrochemical performance distribution, whereas DREAM.3D structures showed poorer performance matches.
Source: §1 · p. 2, §3.3 · p. 15, §4 · p. 15Takeaways & Limitations
The authors validate GAN outputs beyond visual similarity using geometric and topological metrics and simulated electrochemical performance.
Source: §4 · pp. 15–16Obtaining sufficient 3D microstructural training data through experiments or physical simulation can be costly or challenging.
Source: §4 · p. 16
Abstract
from arXiv · showhide
Using a large-scale, experimentally captured 3D microstructure dataset, 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.
1 Introduction
Realistic 3D microstructures are needed for materials simulation, but experimental characterization is costly and many reconstruction methods assume restricted structures. The paper applies a deep-convolutional GAN to complex, multiphase SOFC anodes and reports close agreement with experimental structures and performance distributions.
1 Introduction
- Experimental characterization can limit the size or number of 3D microstructures available for simulation and computational design.The cost or difficulty of characterization motivates synthesizing statistically representative sets of microstructural realizations.Source: §1 · pp. 1–2
- Many established reconstruction methods assume particular structures, limiting their use for heterogeneous, complex material systems.The cited approaches often target binary microstructures, regular geometries, or combinations of the two.Source: §1 · p. 2
- Prior generative-model work included 3D porous media, regular packed grains, and moderate-size three-phase volumes, while much microstructure work focused on 2D images.The paper positions its work as generating complex, heterogeneous 3D structures at a scale suitable for material-property simulation.Source: §1 · p. 2
- For a three-phase commercial SOFC anode, GAN-generated microstructures closely matched the real material’s electrochemical performance distribution in simulations using a locally resolved finite element model.Source: §1 · p. 2
2 Methods
The methods use an experimentally imaged SOFC anode dataset, a grain-based DREAM.3D comparator, and a 3D GAN trained on microstructure volumes. Segmentation and microstructural-property calculations support comparison, while finite element simulations evaluate electrochemical performance.
2.1 Experimental Microstructure Data Set
- The experimental training data are grayscale 3D volumes of interconnected YSZ, nickel, and pores, sampled from a commercial SOFC anode imaged by PFIB-SEM.The volume measures 1697 × 1900 × 124 voxels, or 110 × 124 × 8 µm3, with 65 nm isotropic voxels.Source: §2.1 · p. 3
2.2 DREAM.3D Synthetic Microstructures
- DREAM.3D generated multiphase microstructures by ellipsoid packing using measured phase fractions and log-normal particle-size moments.Forty 12.48 µm cubic volumes were split into octants, yielding 320 subvolumes of 6.24 µm per side.Source: §2.2 · pp. 3–4
2.3 GAN Implementation
- The GAN generator creates candidate images and the discriminator distinguishes generated from experimental volumes through adversarial feedback.Training used batches of sampled experimental subvolumes and generated images; the implementation used upsampling and convolutional generator layers, distributed training on 64 GPUs, and retained the learning rate for stability.Source: §2.3 · pp. 4–5, Figure 2 · p. 5
2.4 Segmentation
- A watershed-based pipeline segmented grayscale volumes into three phases using Sobel gradients, intensity-gradient density maps, manually selected seed bounds, and marker dilation.The bounds were selected using the experimental dataset and reused across 646 subvolumes; the watershed output matched statistically identical results from AVIZO using the same markers. Calculated properties included phase fractions, particle sizes, interfaces, tortuosity, TPB density, and formation factors.Source: §2.4 · pp. 5, 7
2.5 Electrochemistry Simulations
- Electrochemical performance simulations used ERMINE within the MOOSE finite element framework on 30 subvolumes each from the original, GAN, and DREAM.3D sets.A 4.16 µm YSZ electrolyte layer was appended, and simulations covered overpotentials from 0 to 0.4 V.Source: §2.5 · p. 7
3 Results and Discussion
GAN-generated volumes visually resemble experimental microstructures and reproduce their statistical distributions more closely than DREAM.3D; their electrochemical performance distributions also closely match the originals.
3.1 Visual Similarity
- Generated volumes visually resemble the originals, but show less variability, blurrier end slices, and fewer sharp angles in fine features.Source: §3.1 · pp. 7, 9
- Compared with DREAM.3D, GAN morphology more closely resembles the original, with greater phase connectivity and less agglomeration of solid phases.DREAM.3D's ellipsoid-packing scheme constrains clustering and connectivity, which are not directly tunable.Source: §3.1 · p. 9, Figure 4 · p. 10
3.2 Statistical Similarity
- GAN means match original values within one standard deviation, with high overlap across 16 microstructural feature distributions; DREAM.3D shows significant mismatches in metrics (a)–(e).Source: §3.2 · pp. 9, 11, Figure 5 · p. 12
- GAN boxplots overlap the originals around the central distributions, but originals show more outliers spanning a wider range, leaving full variability uncaptured.The authors identify capturing outliers as further work.Source: §3.2 · p. 11, Figure 6 · p. 13
3.3 Electrochemical Performance
- GAN-synthetic and original volumes have highly overlapping electrochemical performance distributions, while DREAM.3D has little overlap at the investigated subvolume sizes.The simulations use meshed microstructures and include only active triple phase boundaries; the authors attribute DREAM.3D's poorer match to differences in active TPB density and local connectivity.Source: §3.3 · pp. 14–15, Figure 7 · p. 15
4 Conclusions
The GAN generated microstructures that closely resembled the experimental structures in visual appearance, geometric and topological properties, and simulated electrochemical performance, outperforming DREAM.3D in structural and electrochemical realism. The work also validates the outputs beyond visual similarity while identifying limits in training data, compute, volume scale, and tunability.
4 Conclusions
- GAN-generated structures closely matched experimental structures across visual, geometric, topological, and electrochemical measures, while DREAM.3D was less structurally and electrochemically realistic.The authors also describe extensive validation beyond subjective visual similarity.Source: §4 · pp. 15–16
- The framework is presented as applicable across materials systems, morphologies, and imaging modes.The authors characterize it as a general and flexible microstructure-generation tool.Source: §4 · p. 16
- The GAN generates realistic structures without user-defined assumptions or prior knowledge of the microstructural metrics it reproduces.Source: §4 · p. 16
- A trained generator can produce arbitrarily many unique volumes, and synthetic microstructure generation is efficient despite substantial training-resource requirements.Source: §4 · p. 16
- Training data can be costly or challenging to obtain, while 3D image generation requires high model complexity and multi-GPU training.Conventional methods such as DREAM.3D can instead generate structures from statistical descriptors without 3D image data.Source: §4 · p. 16
- The implementation generates volumes at 96 × 96 × 96 voxels; larger volumes require proportionally more computational resources.Source: §4 · p. 16
- The present GAN is limited to generating microstructures from the parameter space on which it was trained.Conventional construction methods can vary inputs such as phase fraction and particle size.Source: §4 · p. 16
6 Disclaimer
The disclaimer states that the government, its agencies, and LRST make no warranty or assume liability for the information or disclosed items. It also disclaims government endorsement of named commercial products and says the authors’ views are not necessarily those of the government.
6 Disclaimer
- The government, its agencies, and LRST disclaim warranties and legal liability for the accuracy, completeness, or usefulness of the disclosed information, apparatus, product, or process.Source: §6 · p. 17
- Mention of a commercial product, process, or service does not necessarily imply government endorsement, recommendation, or favoring.Source: §6 · p. 17
- The authors’ views and opinions do not necessarily represent those of the United States Government or its agencies.Source: §6 · p. 17