Predicting energy dissipation in bend flow using explainable artificial intelligence models

Accurate prediction of the energy dissipation rate (η) in bend flows is critical for optimizing hydraulic infrastructure but is often hindered by the high computational cost of 3D Computational Fluid Dynamics (CFD) and the scarcity of experimental data. To overcome these bottlenecks, this study proposes a novel, data-driven surrogate model that integrates Generative Adversarial Networks (GAN) with a Convolutional Neural Network-Back Propagation Neural Network (CNN-BPNN) hybrid architecture. The framework leverages the Design of Experiments (DOE) method to efficiently generate a high-fidelity dataset from validated numerical simulations, incorporating the complex effects of Combined Energy Dissipators of Rough Strips and Guide Piers (RS-GP). To decode the ‘black box’ nature of deep learning, explainable artificial intelligence (XAI) techniques, specifically SHapley Additive exPlanations (SHAP) and Accumulated Local Effects (ALE), are employed to reveal the nonlinear interactions between bend geometric parameters and RS-GP arrangement parameters. Results demonstrate that: (1) The GAN-enhanced CNN-BPNN model achieves superior predictive accuracy (R2 = 0.9650 ± 1.38 × 10-3) compared to traditional machine learning baselines, including BPNN, SVR, XGBoost, KNN, RF, and DT models, effectively serving as a low-cost alternative to extensive CFD simulations. (2) Global sensitivity analysis identifies the relative height of RS (hr), spacing of RS (sr), thickness of GP (td), radius of the bend centre (Rc), bend angle (β), and bottom slope (i) as the dominant factors governing η, establishing an importance ranking of i > Rc > hr > β > sr > td. (3) Significant nonlinear interaction effects are revealed between geometric parameters (e.g. hr-td, hr-Rc, sr-td, td-Rc, td-β), providing new physical insights into how RS-GP configurations alter flow structures to enhance dissipation. (4) Local interpretability analysis offers sample-specific prediction pathways, validating the model's adaptability across diverse flow regimes. This study presents a cost-effective, high-precision tool for hydraulic design, bridging the gap between advanced AI techniques and practical fluid mechanics engineering.HighlightsAn innovative AI-based framework is proposed to effectively predict energy dissipation, significantly reducing the computational cost of 3D CFD simulations.Generative AI-based data augmentation is employed for the first time to address data scarcity in bend flows, validating its feasibility for robust model training.Explainable AI analysis reveals the underlying hydraulic mechanisms and nonlinear interactions between combined energy dissipators of rough strips and guide piers and bend flow dynamics.Local feature attribution validates the model's physical consistency and adaptability across diverse hydraulic operating conditions.The framework serves as a cost-effective, reliable decision-support tool for optimizing flood discharge protection and hydraulic structure design.