Microstructure modeling of deformed alloys using contrastive conditional generative adversarial networks

Published: 08 Oct 2024, Last Modified: 03 Nov 2024AI4Mat-NeurIPS-2024EveryoneRevisionsBibTeXCC BY 4.0
Submission Track: Short Paper
Submission Category: AI-Guided Design + Automated Material Characterization
Keywords: High entropy alloys, contrastive GAN, conditional generation, process-structure linkage
TL;DR: In this work, we propose a contrastive conditional GAN to model the deformation behavior of a high entropy alloy under given processing conditions, as conditional generation of deformed microstructures.
Abstract: High entropy alloys (HEAs) tend to exhibit good mechanical properties, making them potential candidates for various applications. However, tailoring the alloys for target properties requires extensive exploration of microstructure configurations and corresponding properties either through experiments or numerical simulations. Leveraging recent advances in generative modeling, the deformation behavior of CoCrFeNiTa$_{0.395}$ alloy is modeled as conditional generation of deformed microstructures based on processing conditions. To achieve this, a Conditional Generative Adversarial Network (CGAN) model is developed, which synthesizes a deformed microstructure based on temperature and strain rate parameters. A contrastive conditional loss is utilized to induce similarity bias which effectively deals with data sparsity. To help the model learn intricate features across a wide range of process parameters, additional architectural mechanisms like self-attention are employed. Our evaluations reveal good qualitative and quantitative similarities between experimental and predicted microstructures. We also propose a modified contrastive loss for continuous conditioning variables and briefly discuss the ongoing work on demonstrating its generalization capability.
Submission Number: 34
Loading