Development of Process-Structure Linkage Using Conditional Generative Adversarial Networks

Abstract

Abstract: Quantitative characterization of microstructures and understanding of how processing influences these characteristics, subsequently affecting the final properties, are pivotal pursuits for the advancement of materials engineering. Traditional experimental or physics-based numerical simulation methods are resource intensive, limiting their applicability in iterative exploration of large design spaces. To address some of these challenges, in this work, a conditional generative adversarial network (CGAN) is developed for modeling process-structure linkages in a material. Process-structure linkage is posed as a problem of learning the conditional distribution of microstructures given the process parameters and composition. The CGAN model takes composition and process parameters as conditioning variables and transforms the input Gaussian noise (i.e., a sample from the standard Gaussian distribution) into a microstructure (i.e., a sample from the target conditional distribution). The novelty lies in the continuous nature of the conditioning variables, enabling the generation of microstructures for unobserved process parameters and composition. Our findings demonstrate strong qualitative and quantitative resemblance between CGAN-generated microstructures and the reference experimental counterparts. Further, the model exhibits robust generalization to novel process parameters and composition, producing realistic microstructures with features in line with underlying physics. The CGAN model can navigate through the space of microstructures realizable through various initial compositions and processing routes. Consequently, in conjunction with a structure-property linkage model, it offers a potential approach for efficiently selecting compositions and processing routes to achieve desired properties. Graphical Abstract: (Figure presented.). © The Minerals, Metals & Materials Society and ASM International 2024.