Keywords: stress concentration factor, offshore structures, T-joints, finite element analysis, Sobol sequence, Bayesian optimization, neural networks.

A comprehensive database was developed for stress concentration factors (SCF) in offshore tubular T-joints through a code that enables finite element (FE) modeling of a joint using graded mesh generation, load and boundary conditions for a range of geometric parameters. A mesh sensitivity study was conducted and the SCF computations were validated against existing experimental results. A parametric study was conducted to identify the best samples for training a neural network (NN) model. Bayesian optimization by Gaussian Process and Expected Improvement functions were employed to tune the hyper-parameters. A Sobol sampler was used to generate an initial set of points in the search space with the hyper-parameters including learning rate, batch size, number of layers, neurons, activation function and dropout. The optimization process generated a set of trial points using a balanced Sobol sampler, which was evaluated by an objective function that monitored validation loss to obtain the best hyper-parameters. Back-propagation based on a NN model was trained and tested to predict the SCF of T-joints by using the best hyper-parameters obtained from the model. SCF results were compared with parametric equations of Det Norske Veritas (DNV) and Lloyds Register (LR). Advantages of the proposed method have been highlighted.

download the full-text of this paper (PDF, 10 pages, 365 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description