Deep generative learning for microstructure evolution prediction in multimaterial welded interfaces

Abstract

Predicting microstructure evolution in dissimilar metal welds is a challenging task due to complex thermomechanical interactions, heterogeneous material properties, and process-dependent phase transformations. In this work, we present a conditional generative adversarial network (cGAN)-based framework, adapted from image-to-image translation models, for predicting the spatial and temporal evolution of weld microstructures under varying welding conditions. The framework employs a U-Net generator and a PatchGAN discriminator trained using a combined adversarial and reconstruction loss, while being explicitly conditioned on physically meaningful welding parameters such as heat input and cooling rate. Rather than introducing a fundamentally new GAN architecture, the contribution of this study lies in the domain-specific adaptation of the cGAN framework to welding microstructure prediction, including the integration of process parameters as conditioning variables, training on a hybrid dataset combining experimental EBSD (electron backscatter diffraction micrographs) and physics-informed phase-field simulations and extension to both static and time-evolving microstructure sequences. Model performance is evaluated using image similarity metrics (SSIM, PSNR, and RMSE) as well as material-specific measures such as grain size distribution error and phase-region intersection over union (IoU). Results across multiple dissimilar material systems demonstrate that the proposed approach can accurately capture grain morphology, phase boundaries, and temporal evolution trends, highlighting its potential for data-driven weld quality assessment and process optimization.