Generating Synthetic Material Microstructures using Generative Adversarial Networks (GAN)

Synthetically generated microstructure images play a critical role in understanding material behaviors. We investigate a novel AI-driven method for generating realistic synthetic microstructures to enable morphology-controlled computer simulations for material characterization and design.
Generating Synthetic Material Microstructures using Generative Adversarial Networks (GAN)
Like

To read our paper published in Scientific Reports, click here.

Material scientists conduct computer simulations to discover the linkage between microstructural morphology and material properties. In particular, for design and engineering of energetic materials (a term that includes propellants, explosives, and pyrotechnics), computational simulations performed on microscopic images of these materials are necessary to understand the mechano-chemical reactions which can flare up in a few nanoseconds. Energetic materials are typically mixtures of organic crystals, plasticizers, metals, and other inclusions, and these simulations illuminate how their complex micromorphology affects the sensitivity of such materials.

The problem, however, is that current manufacturing processes for energetic material offer little control over the microstructure morphology. This presents a considerable bottleneck to material scientists who desire to precisely control the response of energetic materials to external stimuli; the cut-and-try approach to materials design is extremely expensive and hazardous. Virtually prototyping of energetic materials with precise and predictable sensitivity requires conducting numerous simulations on a large collection of microstructures to accommodate the stochastic nature of the micromorphology. Unfortunately, imaged microstructure samples for a wide class of materials are not in abundant supply.  The research community has found a get-around by developing techniques to create synthetic microstructures; traditionally this has been done by packing geometric primitives (circles, polygons, etc.) that are representative of the real morphology. However, their fidelity to the actual microstructures is questionable as they simplify the morphology; furthermore, packing objects at high densities can be challenging.

Figure 1: A schematic overview of the proposed architecture. The generator takes the local stochasticity parameters ρ and the global morphology parameters λ and produces a realistic synthetic microstructure. The discriminator is then challenged to distinguish synthetic microstructures from real microstructures. The competition between the two results in a generator that produces highly realistic images such that the discriminator can no longer differentiate synthetic images from real images.

In our paper, newly published in Scientific Reports, we propose a novel method for generating synthetic microstructures using generative adversarial networks (GAN). Conceptually, GAN leverages a “competition” of two artificial neural networks, namely a Generator and a Discriminatorthe generator generates a synthetic image while the discriminator attempts to distinguish synthetic images from real images. Interestingly, through such a competition, a generator learns how to generate synthetic images realistic enough to deceive the discriminator. Building upon this concept, we present a novel GAN architecture in our paper, which can generate synthetic microstructure images from parameterized inputs (Figure 1); there is no limit to the packing fraction for solids in our approach. 

Figure 2: Comparison between real microstructures (ground truth) and synthetic microstructures. Both the benchmark transfer learning (TL) method (Li et al.) and the proposed GAN method display realistic image quality overall. However, as seen in the image callouts to the right, small artifacts and blurry boundaries are noticeable in the benchmark method while the GAN method does not produce artifacts.
Figure 3: Distributions of the morphometric parameters indicate the synthetic microstructures are statistically similar to real microstructures. The probability distribution functions (PDF) in each case (real microstructures, TL-synthetic microstructures and GAN-synthetic microstructures) were computed for 25 sample images of size 25μm ×25μm. The curves in different colors indicate the mean PDFs across the 25 images, while the whiskers represent the standard deviations.

Two different types of parameters are used in our GAN model, global morphology parameters λ and local stochasticity parameters, ρ. Global morphology parameters define the overall “style” of the generated image. Local stochasticity parameters add natural stochastic variations to the morphology. Figure 2 shows microstructure images of cyclotetramethylene-tetranitramine (HMX) generated by the GAN model, whose difference from the real microstructure images is not discernable. Morphometry analysis in Figure 3 also suggests the realism of the GAN-generated microstructures. Moreover, this new GAN model can control and manipulate the microstructure morphology. For example, the animation in Figure 4 illustrates how micromorphology changes as the global morphology parameters of GAN are varied smoothly. This capability to smoothly vary the generated microstructures can be used, for example, to control the spatial layout of microscale morphology as seen in Figure 5.

Figure 4: Continuous change of micromorphology achieved via GAN. Intermediate synthetic microstructures are generated by linearly interpolating the global morphology parameters λ.
Figure 5: (a) A graded microstructure generated by linearly interpolating two global morphology parameters. A global morphology parameter λ0 corresponding to small crystal sizes and a parameter λ1 corresponding to large crystals were obtained manually. The image was then generated by spatially grading λ0 and λ1 linearly: λ(x) = (1-x)λ0 + 1, where x∈[0,1] is the horizontal position on the image as a fraction of the image width. (b) A layout of global morphology parameters (left) and the corresponding GAN-synthetic microstructure. Foreground regions were "painted" with λ1 in panel (a) while the background was painted with λ0.

For a materials scientist, our work can be an exciting new tool to conduct morphology-controlled experiments to better understand structure-property linkages. For example, Figure 6 shows an animation of computer simulated shock-induced chemical reactions in two different GAN-controlled microstructuresone with smaller crystals/voids and another with larger crystals/voids. The sensitivity of the two cases are clearly different. For example, the temperature field for the sample with smaller crystals/voids (left column) shows a higher density of hot spots in the control volume when compared to the sample with larger crystals/voids (right column). Also, noticeable difference is seen in the pressure field between the two microstructures; similar to the hot spot temperature field, it is observed that the pqqressure field in the case of the small crystal size sample (left column) achieves higher overall magnitude in a larger part of the domain when compared to the pressure field for the sample with larger crystal. In future work, we will use these new capabilities to conduct controlled experiments and to quantify energetic material response as a function of morphology parameters. This will be significant step forward in energetic material research and will open new possibilities to “design” an energetic material with tailored property/performance characteristics. 

Figure 6: Meso-scale shock simulations performed on a controlled synthetic microstructure containing only small crystals and another have mostly large crystals. Since morphology of the two microstructures are different, reactive behaviors are different: synthetic microstructure with small crystals tends to increase temperature and pressure quicker than synthetic microstructure with large crystals. 

This work was supported by the U.S. Air Force Office of Scientific Research (AFOSR) Multidisciplinary University Research Initiative (MURI) program (Grant No. FA9550-19-1-0318; PM: Dr. Martin Schmidt, Dynamic Materials program). We thank Drs. Chris Molek and Eric Welle from the Air Force Research Lab, Eglin AFB, for providing the images of pressed energetic materials; the images were obtained by Dr. Ryan Wixom at Sandia National Laboratories.

We are always looking for highly motivated and friendly PhD students and postdoctoral scholars eager to work on challenging scientific projects. Interested individuals are strongly encouraged to contact with Prof. Stephen Baek (stephen-baek@uiowa.edu). For more information, please visit our website.

Please sign in or register for FREE

If you are a registered user on Research Communities by Springer Nature, please sign in

Subscribe to the Topic

Electrical and Electronic Engineering
Technology and Engineering > Electrical and Electronic Engineering

Related Collections

With collections, you can get published faster and increase your visibility.

Retinal imaging and diagnostics

This Collection invites works providing insight into using novel or existing retinal imaging technologies in clinical applications or presents new or adaptive forms of these techniques.

Publishing Model: Open Access

Deadline: Apr 30, 2024