Speaker
Description
Microscopic imaging techniques such as computed tomography (CT) are increasingly used tool for the characterization of the microstructure of complex materials because they can provide detailed insight in the materials’ morphology and composition—which is an essential step for deriving microstructure-property relationships. However, a direct microstructural characterization from image data is difficult, as the complexity of discretized image data is often too large, e.g., CT images typically consist of millions of voxels (where voxel refers to the 3D analogon of a pixel). In this talk various applications are presented in which methods from machine learning and stochastic modeling are leveraged for a quantitative microstructural characterization of materials—for more efficient and informative analysis of image data. In the first application, a generative adversarial network (GAN) is deployed to perform super-resolution on scanning electron microscopy (SEM)-images of cycled cathode particles in Li-ion batteries such that fine features like cracks within particles can be more reliably characterized to investigate the state of degradation [1]. Super-resolution of low-resolution images produces highly resolved images, with the same statistically representative field of view that can be produced by high-resolution imaging only with an increased measurement effort. The second application shows how convolutional neural networks (CNNs) can be used to achieve a grain-wise segmentation of 3D image data of polycrystalline materials (such as electrode particles or alloys) [2,3]. Image segmentation is an essential step for the subsequent quantitative analysis of image data. However, image segmentation can be difficult and time-consuming using conventional image processing methods. Therefore, CNNs are becoming increasingly important for image segmentation, as they can be calibrated with relatively small effort to achieve good segmentation results. The third application deals with the structural characterization of materials, using probability distributions of structural descriptors (e.g., size and shape descriptors of grains). Typically, such structural descriptors are correlated with each other which, e.g., may mean that small grains are shaped differently than large grains. Instead of considering distributions of individual structural descriptors (also called univariate or 1D distributions), which do not provide any information on the correlation of descriptors, we obtain a more informative characterization, by modeling the joint probability distribution of multidimensional descriptor vectors [4]. The fourth application deals with an alternative, more advanced characterization method—namely, stochastic geometry modeling [5]. Compared to multivariate distributions (whose realizations are multidimensional structural descriptor vectors that do not necessarily fully describe a material’s microstructure), realizations of stochastic geometry models are virtual microstructures (digital twins) which are statistically similar to the microstructure observed in data—allowing for a holistic characterization of materials. In this application a multiscale stochastic geometry model has been calibrated to image data in order to artificially generate cathode particles with full polycrystalline grain architecture. Besides the holistic characterization of materials, stochastic geometry models allow for the generation of large databases of virtual microstructures with a broad range of structural properties. Such databases can serve as structural input for numerical simulations to subsequently establish microstructure-property relationships (virtual materials testing) [6].
References
[1] O. Furat, D. P. Finegan, Z. Yang, T. Kirstein, K. Smith, V. Schmidt. npj Computational Materials 8 (2022), 68.
[2] O. Furat, M.Y. Wang, M. Neumann, L. Petrich, M. Weber, C.E. Krill III, V. Schmidt. Frontiers in Materials 6 (2019), 145.
[3] O. Furat, D. P. Finegan, D. Diercks, F. Usseglio-Viretta, K. Smith, V. Schmidt. Journal of Power Sources 483 (2021), 229148.
[4] O. Furat, T. Leißner, K. Bachmann, J. Gutzmer, U. A. Peuker, V. Schmidt. Microscopy and Microanalysis 25 (2019), 720-734.
[5] O. Furat, L. Petrich, D. P. Finegan, D. Diercks, F. Usseglio-Viretta, K. Smith, V. Schmidt. npj Computational Materials 7 (2021), 105.
[6] B. Prifling, M. Röding, P. Townsend, M. Neumann, V. Schmidt. Frontiers in Materials 8 (2021), 786502.