Machine Learning-Based Prediction of Hydrodynamic Coefficients and Structural Responses in Tuna Longline Gear

Abstract

The accurate prediction of hydrodynamic characteristics and structural responses in underwater fishing gear is critical for optimizing design, ensuring operational safety, and minimizing environmental impact. To overcome the computational costs and scalability limitations of traditional physical modeling, this study evaluates three machine learning algorithms such as Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Support Vector Machine with a Radial Basis Function kernel (SVM-RBF) to predict the hydrodynamic coefficients and structural responses of tuna longline components, including mainlines and branch lines. Models were trained and validated using a comprehensive flume tank dataset encompassing six gear configurations tested across varying flow velocities and lead-line weights. Results demonstrate that optimal model selection is inherently task dependent. For hydrodynamic coefficients, LightGBM achieved superior predictive accuracy for branch-line drag (whole-dataset R2 = 0.8315), while both LightGBM and SVM-RBF excelled in lift prediction. Conversely, structural responses (sinking depth and x-displacement) proved inherently more difficult to model deterministically due to high-frequency transient dynamics and stochastic variability. While LightGBM provided balanced generalization for sinking depth, SVM-RBF exhibited severe overfitting for x-displacement. In contrast, RF maintained the most conservative and consistent performance across structural targets, effectively mitigating the memorization of dynamic noise observed in the more complex algorithms. Beyond predictive modeling, feature importance analysis identified flow velocity, lead-line weight, material stiffness, and geometric parameters as dominant physical drivers, validating the physical plausibility of the models. Crucially, the integration of experimental and ML analyses revealed that a polylactic acid (PLA)-integrated midsection configuration consistently yielded the lowest and most stable drag force (0.004–0.13 N at 0.49 m/s), representing a 30–60% reduction compared to conventional nylon lines. Furthermore, the study uncovered novel physical phenomena, including velocity-independent deformation stability, progressive transient sinking kinetics, and tension-induced load redistribution. These findings establish machine learning as a reliable, scalable surrogate for longline gear design, advocating for thin-diameter, biodegradable PLA-integrated lines to enhance hydrodynamic efficiency and mitigate marine plastic pollution, while underscoring the necessity of task-specific algorithm selection for robust engineering applications.