GPU-accelerated high-resolution modeling of local scour at river-canal crossings for national water transfer project safety

Accurately simulating the failure mechanisms of protective structures at river-canal crossings under extreme floods urgently demands high-performance simulation tools capable of real-time coupling of the dynamic interactions among flow, sediment, riverbed, and engineering structures. However, existing models are constrained by static boundary assumptions and the traditional two-step approach of flow-sediment simulation followed by post-assessment, making it difficult to dynamically capture the processes of riverbed scour and engineering structure failure. To address this, this study developed a two-dimensional hydro-sediment-morphodynamic model based on an MPI-CUDA heterogeneous parallel architecture. Its core innovation lies in constructing a dynamic boundary feedback framework. By predefining engineering scour-resistance thresholds (such as flow velocity-water depth relationships and design flood peaks) as boundary transition conditions, the model initially realizes a closed-loop simulation logic of the scour-triggered failure – immediate boundary update – dynamic scour response process within numerical simulations. Taking the Yuhe River crossing section of the South-to-North Water Diversion Middle Route (drainage area: 572.7 km2) as a case study, the model integrates unmanned aerial vehicle-derived topographic data with 0.06-meter resolution to construct a high-precision computational mesh comprising 200,000 cells (minimum resolution: 0.5 meters). Leveraging parallel acceleration with four NVIDIA RTX 3090 GPUs, the model achieved efficient simulation of the flood process, realizing a speedup exceeding 800 times compared to traditional serial CPU computation. Validated against the 2021 historical flood event, the model demonstrated a sediment scour/deposition simulation error of less than 9%. Simulation results quantitatively reveal that a 5-year flood (394 m3/s) causes localized failure of gabion baskets, while during a 300-year flood (3591 m3/s), the stilling basin fails but the cutoff wall remains stable. This study not only achieves engineering-scale computational efficiency through GPU heterogeneous parallelism but, more critically, provides an operational preliminary framework for the dynamic coupling simulation of engineering structures and flow-sediment processes.