Generative and explainable artificial intelligence models for enhancing scour depth prediction around a cubical artificial reef
Abstract
Predicting scour around artificial reefs (ARs) is challenging due to the complex interactions between flow properties, sediment characteristics, and AR geometry parameters. This study integrates laboratory experiments with machine learning techniques to predict scour depth around a cubical AR in steady current. A stereo vision system is employed to provide a comprehensive understanding of morphological changes and to accurately measure the equilibrium scour depth around the AR. To address the limitations posed by a small number of physical experiments, a novel Gaussian copula model is proposed to expand the dataset. The performance of three ensemble models (eXtreme Gradient Boosting (XGBoost), Adaptive Boosting, and Random Forest) were compared with traditional models, including Lasso regression and Ridge regression, based on the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and scatter index (SI). The Gaussian copula model significantly enhanced the ML models' performance, with XGBoost achieving the highest accuracy (R2 = 0.95, RMSE = 0.0134, MAE = 0.0118, MAPE = 11.54 %, and SI = 0.1104). Furthermore, an explainable artificial intelligence framework was integrated to assess the influence of various physical parameters on scour depth prediction. This comprehensive approach substantially improves prediction efficiency and provides a viable solution for training models with limited datasets. The findings offer valuable insights for the design of ARs and measures to mitigate scouring.