On the Performance of Physics-Informed Neural Networks for Hemodynamic Predictions in Parameterized Vascular Stenoses

Accurate hemodynamic assessment is essential for characterizing vascular function and pathology. While computational fluid dynamics (CFD) provides the means to simulate blood flow, each anatomical variation requires its own dedicated simulation, which in turn demands substantial computational resources and domain expertise. Accelerating blood flow simulations to enable real-time or near real-time predictions could significantly enhance clinical decision-making and personalized treatment planning. We evaluated single and multi-case physics-informed neural networks (PINNs) in predicting steady-state blood flow in parameterized two-dimensional (2D) stenotic vascular geometries. The PINN was trained without the use of labeled data, utilizing the parameterized incompressible steady state continuity and Navier–Stokes equations. The degree of stenosis was set to vary from 20% to 60% and the Reynolds number (Re) from 500 to 1750. To measure the accuracy, CFD ground truth data were generated using COMSOL Multiphysics® version 6.4. Results show that PINNs accurately predict both axial and vertical velocity fields, with low global and localized errors. Pressure predictions were generally insufficient, particularly in mild to moderate stenoses at low Re, with the median throat-pressure error reaching 38.8%. Pressure is anchored by a single outlet condition, and its non-dimensional scale varies by a factor of 142 across cases, so a few dominate the training objective; hard boundary-condition enforcement improved the field but not the pressure drop. These findings highlight that while current PINNs can reliably reproduce velocity fields, their ability to capture localized pressure dynamics remains limited, indicating the need for more robust formulations.