Optimization of ANN and determination of optimal network parameter range to predict suction caisson foundation capacity

Abstract

Suction caisson foundations are frequently used to moor offshore structures in the oil drilling and wind power generation industries. Though artificial neural network (ANN) models have been successfully applied to predict pile foundation capacity, the considerable differences between the characteristics of pile and suction caisson foundations imply that an ANN model trained using data from the former cannot be applied to predict the capacity of the latter. This study accordingly employed suction caisson foundation data to develop an ANN capable of accurately predicting the capacity of such foundations. The early stopping and model checkpoint techniques were applied to prevent overfitting by saving the immediately prior optimal weight. To obtain the optimal hyperparameter conditions efficiently, a Bayesian optimization algorithm was employed, which significantly reduced the optimization time. This algorithm produced four hyperparameter combinations that exhibited excellent performance; these were each used to train the ANN 500 times, thereby accounting for the uncertainty owing to randomly assigned initial weights. The proposed ANN was subsequently developed using two approaches: parameter analysis and optimization. The parameter analysis determined that the optimal number of network parameters for the selected hyperparameter combinations was 7,638, which was within the 500-650,000 range determined by a general analysis. The verification root mean square error(RMSE) of the ANN model developed using the optimization process was 8.88 with a coefficient of determination of 0.9998. Notably, because suction caisson foundation data have characteristics consistent with general geotechnical engineering practices, the optimal network parameter range and optimization method employed in this study to develop the ANN can be used with other data obtained in the geotechnical field.