A Generative Adversarial Network Approach with a Random Patch Discriminator to Generate 3D Synthetic Microstructures Containing Second Phase Particles

Abstract

The failure of metals under quasi-static and dynamic loads is influenced by second-phase particles that act as failure nucleation sites. It is desirable to have a large quantity of data when developing models to understand how particles contribute to failure, but collection of such data from real samples is prohibitively costly and time consuming. In this work, a machine learning approach is developed to learn the features of a microcomputed tomography dataset and output 3D synthetic particle microstructure volumes. The machine learning approach comprises a generative adversarial network in conjunction with the random patch discriminator method. The method maintains training stability when generating volumes up to 128×128×128 voxels by using a random patch method for the discriminator model. This random patch method takes advantage of the ability to sample smaller sections of the whole data to determine whether characteristics within the samples are representative of the microstructure. Generated volumes are compared against the real dataset to demonstrate the ability to capture all features of the sparse particles when trained from a relatively small set of data. Statistical accuracy is verified by comparing the distribution of volume, size, and shape of the particles, as well as two-point and clustering correlation functions. We show that our machine learning approach is particularly well suited for describing two-phase data with disparate phase fractions and data that possess anisotropy in the size, shape, and spacing of individual phases compared to other approaches that describe two-phase materials with similar volume fractions and more isotropic features.