Artificial neural network prediction of buckling load of thin cylindrical shells under axial compression. (1st December 2017)

• Record Type: Journal Article
• Authors: Tahir, Zia ul Rehman; Mandal, Parthasarathi
• Is Part Of: Engineering structures. Volume 152(2017:Dec. 01)
• Page Start: 843
• Page End: 855
• Publication Date: 2017-12-01
• Subjects: Thin cylindrical shells -- Classical buckling theory -- Knockdown factors -- Artificial neural networks -- Bayesian regularisation backpropagation; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2017.09.016
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: ANNs using Bayesian regularisation backpropagation used to predict buckling load. 370 shell specimens were used to train and 20 specimens to validate the ANN models. The predicted buckling load were compared with NASA and EC3 design recommendations. The ANN models predict buckling load within 10% of the experimental buckling load. NASA SP-8007 and EC3 provide 10–50% conservative estimates for buckling loads. Abstract: Thin-walled circular cylindrical shells under axial compression are prone to buckling; the reduction of buckling load from the theoretical estimation is considered primarily due to imperfection sensitivity. The buckling load from carefully conducted experiments using nominally similar shells falls below the prediction by the classical theory with substantial scatter. The current design recommendations apply highly conservative knockdown factors to the theoretical buckling loads to estimate the load carrying capacity of the shell structures. In this study, a systematic analysis of experimental data from the literature has been conducted using the artificial neural network (ANN). The networks were trained using Bayesian regularisation backpropagation training function. Two network models with eight and ten neurones were used to train, test and validate 390 sets of experimental data. The buckling loads predicted by the ANN models were compared with the design recommendations by National Aeronautics and Space Administration (NASA), Eurocode 3 (EC3) and the experimental buckling loads. The ANN models predict buckling load within 10% of the experimental buckling load and can be reliably used within the parametric range used in training. The NASA design recommendations provides 10–50% conservative estimates compared to the experimental loads while EC3 predictions are conservative by more than 50%.