Deep-Neural-Network-Based Optimal Design of Cylinder Structures Under Hydrostatic Pressure

The structural design of unstiffened cylindrical shells under external hydrostatic pressure is critical for the safety of marine structures, such as submarine hulls and pressure vessels. Accurately assessing nonlinear buckling and collapse failure modes traditionally requires computationally intensive Finite Element Analysis (FEA), which creates a bottleneck in iterative design optimization. To address this, our research leverages a robust Deep Neural Network (DNN) model specifically trained and validated for AL-6061 aluminum alloy cylinders. This predictive model, focusing on unstiffened cylindrical shells within a valid domain (2≤L/D≤15 and 20≤D/t≤150), integrates a high-speed surrogate model with a Differential Evolution (DE) algorithm. This predictive model was trained on a large-scale dataset of 46,060 points generated through FEA simulations and rigorously validated against 28 physical experimental data points. Building upon this foundation, the present study implements a novel optimization framework that integrates the pre-trained DNN as a high-speed surrogate model with a Differential Evolution (DE) algorithm for global optimization. The primary objective is to minimize structural weight while strictly satisfying collapse strength requirements. Additionally, a grid search component is incorporated to provide designers with multiple feasible design candidates almost instantaneously. Validation against independent FEA results confirms high fidelity, with error rates of less than 2%. This methodology transforms the design cycle from days to mere minutes, establishing a reusable digital asset that significantly enhances efficiency and structural safety in marine engineering.