Machine learning applications to load and resistance factors calibration for stability design of caisson breakwater foundations

Due to the limit state functions commonly defined in implicit fashions, calibrations of load and resistance factors for limit state designs of breakwater foundations using Monte Carlo simulations (MCSs) are time-consuming and computationally expensive. This study proposed a practical framework combining the newly developed metamodels and an efficient optimization to address these computational issues. For this purpose, two metamodels of artificial neural network and Gaussian process are constructed to replace actual implicit problems. Two approaches to developing the metamodels, where the limit state data (LS data) is included in the training data or not, are investigated. Furthermore, an indicator that helps to stop improving metamodels is integrated into the optimization process to avoid redundant calculations effectively. Finally, the efficiency of the proposed framework and the accuracy of the newly developed metamodels is validated with basic MCSs using two case studies of breakwaters. The results indicated that using LS data in the training dataset helps to construct metamodels quickly and accurately. Namely, the necessary training data can decrease by about half compared to the conventional approaches, regardless of the metamodels used. Remarkably, the proposed framework can dramatically shorten the computing time from days (when using the basic MCS) to tens of minutes.