Abstract
Finite element modeling (FEM) is the standard way to evaluate structural responses based on the physics of the application. They are often used for time-domain analysis, simulating a set of stochastic loading cases on slender marine structures, and carrying out different types of nonlinearities. While the irregular wave approach offers a closer approximation to structural responses due to environmental loadings’ stochastic nature, it is more demanding computationally. To reduce simulation time while ensuring accuracy, this study introduces strategies to design artificial neural network (ANN) frameworks using deep learning algorithms. The LengthNet class type can predict responses at multiple structural nodes simultaneously, both temporally and spatially, cutting down training time, especially when dealing with a high number of nodes. This work introduces the LengthNet model based on causal convolutional neural networks, which outperforms the state-of-the-art LengthNet based on Recurrent Neural Networks and the NodeNet model. A significant improvement in the PAC (Percentage of Accurate Cases) metric was observed for the Lazy-Wave configuration. Different fatigue regions are processed at the same time, going from the upper regions down to the touchdown zone (TDZ), passing through local analysis and fatigue quantification via the S-N curve. Predicted damage values versus the ground truth ones are also presented.
