Background: The lower Taedong River (DPR Korea) is a typical fluvial-estuarine-marine composite sedimentary zone characterized by strong heterogeneity, complex hydrodynamics, and highly variable sediment properties. These geological and hydrological conditions give rise to severe drilling challenges, including unstable penetration rates, low construction efficiency, frequent bit clogging, and borehole collapse. This case study proposes a deep learning-based solution to address these engineering problems.
Case Presentation: A total of 350 field datasets were collected from 50 drilling points in the study area, encompassing 11 input parameters classified into three categories: 5 sediment geological indices, 3 hydrodynamic parameters, and 3 drilling operational parameters. A 4-layer Backpropagation (BP) neural network with a topological structure of 11–15–10–1 was constructed. The Adam optimization algorithm and early stopping strategy were employed in model training to mitigate overfitting.
Results: The trained BP neural network model achieved excellent prediction performance on the test set, with a root mean square error (RMSE) of 0.12 m/h, a coefficient of determination (R²) of 0.97, and an average relative error of 2.8%. This performance is significantly superior to that of the multiple nonlinear regression model (RMSE = 0.16 m/h, R² = 0.87, average relative error = 6.8%). More than 92% of the model’s predictions had a relative error within ±5%, meeting the accuracy requirements of practical geotechnical engineering.
Conclusion: This case study validates the effectiveness of deep learning in accurately predicting drilling rates in complex composite sedimentary environments. The proposed BP neural network model serves as a practical engineering tool for optimizing drilling parameters, reducing construction costs, and enhancing safety control in similar coastal drilling projects worldwide.