Residual U-Net for accurate and efficient prediction of hemodynamics in two-dimensional asymmetric stenosis

This study presents residual U-Net (U-ResNet), a deep learning surrogate for predicting steady hemodynamic fields in 2D asymmetric stenotic channels at Reynolds numbers 200-800. By integrating residual connections with multi-scale feature extraction, U-ResNet achieves high accuracy while reducing computational costs compared to computational fluid dynamics (CFD). Evaluation against U-Net, Fourier neural operator (FNO), and U-Net enhanced FNO (UFNO) shows U-ResNet’s superior ability to capture sharp hemodynamic gradients and complex flow features. For pressure, wall shear stress, velocity, and vorticity predictions, U-ResNet consistently achieves lower normalized mean absolute errors (1.10%, 0.56%, 1.06%, and 0.69%, respectively) than other models. It also generalizes robustly to unseen Reynolds numbers without retraining (rare among existing models). Computationally, U-ResNet offers a 180-fold speedup over CFD, cutting simulation time from ~30 minutes to 10 seconds. Its non-dimensional design ensures scalability across vessel sizes and anatomical variations, broadening clinical applicability. U-ResNet’s narrower error distributions further confirm its reliability. These strengths make U-ResNet a promising tool to complement CFD for real-time decision support, treatment planning, and device optimization. Future work will extend to 3D geometries and patient-specific data.