Predicting Post-critical Load Drop in Conical Shells Through Artificial Neural Network

Thin truncated conical shells are widely employed as adapters between cylindrical shells of varying sizes in launching type vehicle systems. The conical shell subjected to axial compression is reported to be highly imperfection sensitive, leading to a significant disparity between the theoretical and the experimentally obtained critical load. NASA SP-8019 stipulates a single Knockdown Factor (KDF) value of 0.33 across all type of conical shell geometries for designing which is a highly conservative estimate. The study employs an existing experimental database, augmented with more data from the Finite Element (FE) based nonlinear stability analysis to accurately predict the post-critical drop through KDFs. The prediction uses Artificial Neural Networks (ANN) for training of the available data set. The entire data is segmented into two bins, namely the training dataset (90% of the data points) and the testing dataset (10%). The process of training involves an algorithm based on the Bayesian regularization back propagation. The ANN models are tested and validated with different numbers of neurons. The ANN configuration adopted in the study predict the critical load/KDFs with higher accuracy. KDFs predicted by such ANN are also less conservative than the other available code stipulated KDFs.