A method for predicting the hydrodynamic characteristics of trawl codends based on machine learning models

• This study evaluates key fishing design parameters to predict the trawl codend drag and dynamic motions using five machine learning algorithms, establishing a data-driven optimization framework. • Mesh size, twine thickness, and codend configurations affect hydrodynamics; T90 mesh reduces fluttering motion, enhancing selectivity. • Identifies GBGAM and GPR as the best stable models, hence expediting the iteration of fishing gear design through accurate hydrodynamic predictions. • Drag oscillations are dominated by low-frequency structural dynamics, revealing ways to control codend motion and reduce non-target catch. • The proposed machine learning approach substitutes conventional physical modeling, facilitating accurate trawl performance simulation and offering actionable insights for intelligent, sustainable trawling. Accurately estimating the hydrodynamic behavior of a trawl codend is crucial for optimizing fishing gear selectivity, minimizing non-target fish captures, and protecting biodiversity in sustainable fisheries. This study evaluates and predicts the influence of key design parameters (mesh size, twine thickness, twine material, mesh orientation, knot type, codend length, liner, and mesh number), as well as current speed, on both drag force and fluttering displacement of the codend. Five machine learning models, namely the Multilayer Perceptron (MLP) neural network, Random Forest (RF), Support Vector Machine (SVM), Gaussian Process Regression (GPR), and Gradient-boosted generalized additive model (GBGAM), were employed and rigorously validated using K-fold cross-validation to establish a comprehensive, data-driven framework for codend performance optimization. The results indicate that the mesh size, twine thickness, and codend configuration exert a strong influence on codend drag and dynamic motion, with T90 codends exhibiting lower fluttering and higher drag than T0 codends, suggesting that replacing T0 with T90 may improve size selectivity by stabilizing codend behavior. Furthermore, the drag force oscillations are directly linked to codend dynamic motion, driven predominantly by low-frequency structural dynamics and modulated by design factors and current speed. Regarding the predictive performance, GBGAM ( R 2 = 0.996, MAE = 0.8, and RMSE = 1.18), GPR ( R 2 = 0.995, MAE = 0.73, and RMSE = 0.91), and MLP ( R 2 = 0.989, MAE = 0.96, and RMSE = 1.21) outperform the other models in predicting drag force. For codend displacement, GBGAM ( R 2 = 0.9345, MAE = 0.35, and RMSE = 0.50) and GPR ( R 2 = 0.841, MAE = 0.45, and RMSE = 0.60) exhibit robust predictive capability. Selecting the most important features for trawl codend performance reduces computational time while improving model accuracy. Both GBGAM and GPR exhibit strong stability and generalization capacity, as indicated by their consistent hydrodynamic force and codend displacement predictions during 10-fold cross-validation. The proposed machine learning framework thus provides a powerful alternative to traditional physical modeling methods, enabling precise simulation of trawl performance and yielding actionable insights for the advancement of intelligent trawl nets.