Towards efficient uncertainty quantification with high-resolution morphodynamic models: A multifidelity approach applied to channel sedimentation

Abstract To guarantee port accessibility, navigation approach channels need to be well maintained. Annual dredging efforts to maintain navigable channels may well exceed tens of millions m 3 of sediment per year, which results in high recurrent costs for port operators. Quantification of expected siltation rates with process-based numerical models helps to effectively design and optimise approach channels. The setup of such models requires several assumptions and parameter settings which introduce uncertainty in model output. However, traditional Monte Carlo methods to quantify that uncertainty in model output are often too resource-intensive with current standard computer resources to be feasibly applied in coastal engineering projects such as approach channel design. Here, we use an alternative multifidelity approach to estimate the probability density function of channel siltation, at lower computational costs compared to direct Monte Carlo simulation. The idea behind this method is to map the output uncertainty of a faster, but inaccurate model to a preferred high-detailed model. The key requirement is that the faster, low-fidelity model and the detailed high-fidelity model are correlated, and that his correlation can be modelled with a probabilistic function. Since linearity of the correlation is not a requirement, the coarse-grid model can be very inaccurate but still serve as an adequate predictor of the high-fidelity model. In this study we did observe a highly nonlinear correlation, which in our case is explained by underestimation of channel siltation near the surf zone by the coarse model. In the presented multifidelity approach we adopted a combination of quasi-random Monte Carlo simulation and a non-parametric Gaussian process transfer function to estimate the uncertainty of total siltation and spatial patterns of siltation in a port approach channel. We argue that the multifidelity approach is conceptually straightforward and found that it can be used to significantly decrease the costs of probabilistic analysis; in our case we found a 85% decrease compared to direct Monte Carlo simulation. An additional advantage is that the approach allows for a trade-off between precision and efficiency by varying the number of high-fidelity model runs. Therefore, we conclude that the multifidelity framework is a potential powerful alternative for cases in which direct Monte Carlo simulation is infeasible or undesirable.