Offshore turbidity currents forecasting (part I): Integrating deep learning and computational fluid dynamics

Abstract

This study introduces a novel methodology for real-time forecasting of turbidity currents, a significant geohazard to offshore infrastructures such as pipelines and submarine telecommunication networks. Numerical modeling, deeap learning (DL) techniques, and field measurements were incorporated to provide accurate predictions of upcoming turbidity current events. A forecasting model was developed using a combination of two DL methods, Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs). This combination is helpful in capturing both spatial features and temporal sequences in the data, making it well-suited for predicting the movement and behavior of turbidity currents. The DL models (CNN-LSTM) were initially trained using synthetic data from Computational Fluid Dynamics (CFD) analyses, and then real field dataset from an Acoustic Doppler Current Profiler (ADCP) in the Congo Canyon, West Africa. The CFD analysis incorporated a turbulence model known as the renormalization group k-epsilon model to enhance accuracy, and it was validated with an experimental dataset. To fine-tune the DL models, Bayesian hyperparameter tuning was employed, a method that systematically adjusts the settings of the model to enhance performance. Also, employing transfer learning, a technique that transfers knowledge from simulated data to actual field data, reduced prediction errors by 50 %. The pretrained models were fine-tuned and the impacts of the key hyperparameters such as input-label width ratio, forecast window length, and data resolution were investigated. After fine-tuning, the optimized models achieved prediction accuracies with less than 10 % error, utilizing data sampled every 480 s across a forecasting window of 40 h. Overall, the proposed DL methodology presents a promising basis for an AI-based early warning system against turbidity current hazards for offshore and marine infrastructures.