Fast Prediction of Dynamic Structural Response Using Reduced Basis Function Combined with Neural Network

Abstract

Frequent calibration of the dynamic characteristics of jack-up platforms are important for accurate evaluation of their safety and comfort to occupants. As jack-ups are large in scale involving many physical mechanisms and hydrodynamic coupling, it is challenging to rapidly optimize the estimation of system parameters and predict the dynamic responses under potential scenarios. This study develops a novel offline-online framework to address these issues. In the offline phase, a set of reduced basis functions is extracted from a collection of high-fidelity datasets, and the corresponding coefficients are employed to train neural network models. The online phase involves mapping model parameters to the coefficients of reduced basis functions using the trained neural network. The trained neural network combined with the reduced basis function is then utilized to predict dynamic response. The feasibility of the proposed method was evaluated through a numerical model of a jack-up under wave loads. The results indicate that the method is effective, robust, and promising for the rapid evaluation of large-scale structures.