Deep Learning Design of Graphene-Reinforced Polyurethane Foams from SEM Microstructure Images and Style-based Generative Adversarial Networks

Polyurethane (PU) foams are a well-known class of polymers with a variety of industrial applications. Novel foam designs can be obtained by altering various properties such as mechanical, electrical, thermal properties which are often dependent on microstructure characteristics (cell size, density, etc.). Artificial intelligence techniques have been rapidly progressing in the past decade and have become an important tool for efficient material discovery enabling high throughput materials characterization and structure-property modelling [1]. Synthetic micro-structure images have significant potential to accelerate materials discovery. However, there are several limitations of current methods in evaluating generative models from materials science domains such as small dataset sizes, images with features different from those in ImageNet, and the need to score based on physically accurate microstructure characteristics that can be synthesized in the lab. Unlike well-studied datasets (e.g. ImageNet), many of the features of a realistic micro-structure image cannot be recognized by the human eye, thus limiting the applications of a model pretrained on ImageNet as the base of a material’s synthetic microstructure evaluation. Beyond image quality and diversity, in the material science domain, such method should generate microstructures which can be synthesized in an experimental setting. We discuss a case study where we demonstrate the use of deep learning techniques for modeling and optimizing the design of graphene-reinforced PU foams. In particular, we leverage state of the art techniques in generative design and latent space interpolation [3,4] to generate synthetic images with desired properties. Graphene-reinforced PU foams were synthesized with varied amounts of additive (Control, 0.01%, 0.025%, 0.05%) and mechanical properties were measured with Instron machine ASTM 3574-08 following Test C, E, Die C while grayscale SEM images were taken using SE2 secondary electron on JEOL JSM-6610 SEM. We apply inverse projection using gradient descent to match a training image to corresponding latent vector in the trained StyleGAN2 network generating interpolations that are more realistic based on learned representations of the latent vectors, see Fig. 1. The mechanical properties of PU foams can vary with changes in foam cell microstructure such as cell size, density, surface morphology and texture which in turn are correlate to their physical properties. A deep learning-based image processing technique was used to perform automated labeling and image segmentation on the PU foam SEM images as in Fig. 2 (a). The predicted properties (using surrogate deep-learning models) show desired variability in resulting synthetic images predicted mechanical properties, see Fig. 2. (b). Further, microstructure-mechanical performance correlations as well as new insights into composition/structure/property/performance relations can be uncovered [5, 6] from virtual latent space interpolation experiments by varying the synthesis/processing parameters. In conclusion, we demonstrate that state-of-the-art deep learning techniques can be leveraged to characterize and efficiently design graphene-reinforced polyurethane foams with targeted properties. Synthetic images generated using StyleGAN2+DiffAug [2] through latent space interpolation on SEM microstructure data of PU foams with varying mass % of materials A and B. (a) Deep learning-based image processing allows to perform automated labeling and image segmentation (b) Using two synthetic images, linear interpolations in the latent space shows linear progression of the compressive modulus as verified by an ML based surrogate prediction model.