Stress field prediction: UNet model integrated with FocalNet Transformer

Abstract

Design generation and optimization using state-of-the-art deep learning (DL) techniques, especially convolutional neural networks (CNNs), has become a cornerstone in engineering in today's paradigm. Traditional methods like Finite Element Analysis are computational and time intensive. Ensuring structural integrity presents a significant challenge, which has prompted engineers to shift towards data-driven approaches in design engineering. This study presents a novel method that combines CNN and transformer models to predict stress fields in industrial designs, thereby increasing structural reliability. Transformer modules originally developed for integrating natural language processing tasks reveal a versatile and scalable DL system in this domain. By combining hierarchical representations with fewer transformers and the spatial understanding of CNN, engineers can achieve unprecedented accuracy in stress calculations while maintaining structural robustness. The proposed model leverages the UNet model, and CNN structure combined with the transformer model, i.e. FocalNet Module for prediction. The model is trained on a dataset of stress-annotated data sets

this model learns the complex relationship between design features and structural stress, resulting in a robust stress prediction tool. Blending the CNN-based UNET model with transformer architecture has profound implications, especially in industries that require rapid generation of conventional designs, such as aerospace and automotive engineering and enhance safety standards That research presents a novel approach that combines CNN-based stress estimation with transformer modules, demonstrating its potential to revolutionize design optimization in various engineering fields.