Prediction of short-term non-linear response using screening combined with multi-fidelity Gaussian Process Regression

Abstract

Predicting wave impact design loads is crucial for ensuring safety and performance of maritime structures, but it is challenging due to the complexity and rarity of these events. Existing methods are mainly suitable for prediction of weakly non-linear responses, or are very computationally expensive. Highly nonlinear responses require a fidelity level that can only be achieved with expensive CFD or experiments, leading to sparsely populated exceedance distributions. A new event-based multi-fidelity method called ‘adaptive screening’ therefore combines elements of screening, multi-fidelity Gaussian Process Regression and adaptive sampling, to more efficiently predict highly non-linear loads. It is applied at the level of the response peak exceedance probability distributions. A simplified case study using second-order wave data validates the effectiveness of the method in accurately predicting short-term design loads. The new method predicts more accurate MPM results than the conventional method recommended by class societies and the ITTC, while also significantly reducing the required HF simulation time. The new method has a deviation of only 0.3–3.5% from the true 1-hour MPM over all test cases, compared to the conventional method’s deviation of 5.2–6.7%. The HF simulation time required to do this is 91 times shorter with the new method (0.033 versus 3 hours per sea state). The new method is not very sensitive to input noise as long as HF samples are selected properly, and the application of the method to the exceedance distributions works.