Comparison of deep learning techniques for prediction of stress distribution in stiffened panels

• CNN and MLP are used for prediction of von Mises stress across stiffened panels • Approach to encode geometry, loads, and BCs for CNN input and output is proposed • CNNs predict 3D stress distributions in stiffened panels with <5% error vs. FEM • Skip connections (U-Net) enhance CNN performance for complex structural analysis • PCA reduces MLP training complexity but can only be used for simpler cases Compared to the finite element method (FEM), surrogate models for structural analysis enable more efficient assessment of a response under loads and subsequent optimization. Recent advancements in deep learning have allowed use of neural networks as surrogate models in a variety of fields with astonishing results. Nonetheless, their use for predicting stress distribution in stiffened panels is unexplored. Predicting stress fields is important for various limit states. We propose an approach to encode stiffened panels with various geometries into grid spaces, which can then be processed by a convolutional neural network (CNN). Uniform pressure and patch loading are considered. The performance of CNN using the proposed modeling approach is compared to multilayer perceptron (MLP) in predicting von Mises stress distribution in stiffened panels. Principle component analysis (PCA) is used to reduce the training complexity for MLP. Moreover, the effect of skip connections is investigated utilizing two different CNN architectures. Five case studies are conducted to assess the performance of these neural networks in predicting the stress distribution in stiffened panels across various geometric configurations, including variations in the number of stiffeners, loading and boundary conditions. The study reveals that CNNs, particularly with skip connections (U-Net), outperform MLP, achieving less than 5% mean absolute percentage error with respect to FEM results in all cases. MLP with PCA achieves satisfactory results for simpler problems, but cannot be trained for more complex tasks. CNNs effectively capture local stress variations in all cases. CNNs have good capability in predicting stress distribution with limited amount of data, making them a viable tool for real-world structural analysis.