Reliability consistent knockdown factors for truncated conical shells via artificial neural network (ANN) based predictions and meta-modelling

• Published in: Thin-walled structures 2024-12, Vol.205, p.112541, Article 112541
• Main Authors: Majumder, Rohan, Mishra, Sudib Kumar
• Format: Article
• Language: English
• Subjects: Artificial neural network (ANN); Conical shell; Knockdown factor (KDF); Reliability based design (RBD); Testing; Training; Validation
• ISSN: 0263-8231
• DOI: 10.1016/j.tws.2024.112541

•KDFs for Truncated Conical Shells under axial compression is predicted through ANN.•A feedforward-backpropagation ANN with fifteen input parameters are used in prediction.•Extensive experimental and simulation datasets are used for training, testing and validation.•ANN offers precise, less-conservative KDFs than the code(s) using only two/three inputs.•A reliability consistent design for the KDFs is proposed via ANN based metamodeling. Utility of conical shells ranges from aerospace launch vehicles to atomic force microscopy. Thin-walled conical shells are susceptible to buckling under axial compression. Due to remarkable disparities between the experimental and theoretical critical load, the actual critical load is obtained by multiplying the theoretical critical load with a highly conservative Knock Down Factor (KDF). Such conservatism may be relaxed for economic design. This being the eventual goal, the Artificial Neural Network (ANN) is employed herein for an improved prediction of the KDFs using experimental data on Mylar cones (non-metal) and Finite Element (FE) simulation data from metallic cones. The ANN uses fifteen input parameters for describing the conical shell to train the network with the KDFs being the sole output. The datasets are segregated into the training and testing (T&T) and the validation dataset. The latter demonstrates that the prediction errors mostly lie within ±5 percent. The code stipulated KDFs (NASA, Euro code EC-3) are compared with the experimental, simulated and the ANN predicted KDFs. The ANN based KDFs are shown to be highly accurate, less conservative and more precise than the existing recommendations, partly due to a wider set of input parameters (than the code provisions) used for the network training and predictions. Furthermore, the ANN predicted KDFs are employed to conduct a reliability-based design (RBD) of truncated conical shells having random geometric imperfections. The RBD furnishes the pertinent design variables (i.e., semi-vertex angle θ) for a target reliability index (β‾) and respective KDFs for varying imperfections and shell geometries.