Conditional generative adversarial network for welding deformation field prediction of butt-welded plates

Abstract

Predicting deformation field of butt-welded plates induced in welding process is a beneficial task to improve the construction quality of ships and offshore structures. However, traditional numerical approach and experimental test could not predict the deformation at small costs of computing resources and time. To solve this problem, an advanced surrogate model of conditional generative adversarial network (cGAN) is developed in this paper to carry out the forecast task quickly. Two competitors of a generator and a discriminator contained in the network are optimized through a series of adversarial training iterations. Welding parameters and generated deformation are collected from a number of numerical welding simulations of a butt-welded plate with varying thickness and width. These variables are coded into independent color images using mapping algorithm to train the network. The deformation field obtained via the surrogate model and numerical method are compared based on their relative difference of color values. Comparison demonstrates that the two results agree well and the robustness of the proposed predictive model can be verified.