Rapid Prediction of Three-Dimensional Scour Flow around Bridge Piers via Body-Fitted Coordinate-Based U-Net

TL;DR

This paper introduces BFC-UNet, a physics-aware deep learning model that rapidly predicts 3D turbulent flow around bridge piers, enabling real-time risk assessment and digital twin applications by accurately reconstructing complex flow features with high speed.

Contribution

The study develops a novel BFC-UNet architecture that preserves geometric integrity in 3D flow predictions, significantly improving speed and accuracy over traditional CFD methods.

Findings

Predicts velocity, pressure, and shear stress with R2 > 0.98.

Infers full 3D domain in 8 milliseconds on GPU.

Captures vortex topology and flow evolution accurately.

Abstract

Predicting three-dimensional (3D) turbulent flows around bridge piers is a prerequisite for assessing local scour, a primary cause of infrastructure failure. While Computational Fluid Dynamics (CFD) captures complex flow features - such as horseshoe vortices - its high cost hinders real-time risk assessment. This study presents a physics-aware deep learning surrogate using a Body-Fitted Coordinate (BFC) system, BFC-UNet, designed to rapidly reconstruct 3D Reynolds-Averaged Navier-Stokes (RANS) solutions on curved domains. Unlike voxel-based Convolutional Neural Networks (CNNs) prone to staircase errors, the proposed architecture leverages a BFC system to predict the bed shear stress accurately. By transforming the physical O-grid into a canonical computational space, the model preserves the geometric integrity of the no-slip boundary. Trained on 2,304 simulations parameterized by inlet velocity and scour depth, BFC-UNet predicts velocity, pressure, and bed shear stress distributions with an R2 value of > 0.98. It infers a full 3D domain (200,000 cells) in just 8 milliseconds on a single Graphics Processing Unit (GPU) - achieving a speed-up of five orders of magnitude over the CFD solver. Crucially, the model captures the topological evolution of vortex structures, including wake expansion and diving flows. These findings position BFC-UNet as a promising foundation for real-time digital twins, bridging rigorous fluid mechanics with data-driven efficiency.