Deep learning for predicting welding distortions in T-joints

Abstract

Stiffened panels are common structural units in marine vessels, civil and aerospace structures. Production via welding can lead to excessive distortion of their plates, which negatively affects structural integrity and dimensional accuracy. Conventional practical approaches for distortion control are costly, making simulation models attractive tools to mitigate distortions. Prediction of welding distortions using finite element (FE) simulations is computationally intensive, especially when used repeatedly in design and optimization. Effective surrogate models in the form of deep learning could alleviate this issue, but the selection and construction of deep neural networks for this purpose are presently unclear. This study focuses on predicting welding-induced distortions in a T-joint. Two neural networks—a multilayer perceptron (MLP) and a convolutional neural network (CNN)—are employed to predict distortions. Two case studies are conducted for each model, exploring variations in geometry and welding sequences. The database is generated from FE simulations of the gas metal arc welding (GMAW) process. The effects of welding order and direction on distortions are studied, concluding that an appropriate selection of welding sequence and direction can reduce distortion by up to 40%. This data is then used to train the neural networks. The MLP and CNN models are designed and trained to predict distortion fields by tuning their architecture and other hyperparameters. Results demonstrate that both models are effective; however, the CNN achieves higher accuracy for complex distortion patterns, highlighting its suitability for more intricate scenarios.