(V) Generative Adversarial neural Networks (GANs) for 3D microstructural design and optimization

Collecting ensembles of statistically equivalent, realistic, 3D microstructures is a critical component of the ICME pipeline. However, the number of data sets that can be directly obtained by 3D microscopy, such as synchrotron methods (HEDM, 3DXRD) and serial sectioning (RoboMet, plasma FIB), remains limited due to the cost and accessibility of these techniques. Therefore, the sparse experimental data is often supplemented by additional structures generated computationally. While most synthetic microstructures are constructed from physical or geometric models, machine learning offers a new approach for learning and reproducing the the underlying geometry and topology of complex microstructures. In this work, based on the convolutional neural network (CNN) structure, a class of generative models termed Generative Adversarial Networks (GANs) has been implemented to learn and generate 3D microstructures of solid oxide fuel cell electrodes. Besides visual similarity, the generated and the original microstructures have measurably similar 3D topological properties such as surface area, triple-phase-boundary density, volume fraction, and particle size. Furthermore, we show that, via simulation, the electrochemical properties of the synthetic microstructures match the properties of the experimental system. There is profound significance in the ability of generative models to recreate microstructures with high fidelity; the essence of complex and three-dimensional microstructures may be captured and represented in a compact form that can be used to generate a nearly limitless set of related structures.