Near-wall Blood Flow Modeling with Physics-Informed Neural Network (PINN).

16th U.S. National Congress on Computational Mechanics (USNCCM) conference presentation.

Title: Hybrid Physics-based and Data-driven Modeling of Near-wall Blood Flow with Physics-Informed Neural Networks (PINN)

Keywords: Scientific Machine learning; Hemodynamics; Wall shear stress; Sparse data; Data-driven modeling.

Abstract:
Near-wall blood flow and wall shear stress (WSS) regulate cardiovascular disease, yet they are challenging to quantify with high fidelity. Computational fluid dynamics (CFD) models can leverage high-performance computing to enable high-resolution and accurate quantification of WSS. However, these models suffer from uncertainty in parameters and boundary conditions. On the other hand, direct experimental measurement in-vivo could reduce uncertainty, however, these measurements typically do not meet the spatial resolution requirement in calculating WSS and struggle to accurately quantify near-wall flow. Physics-informed neural networks (PINN) provide a flexible machine learning framework to integrate mathematical equations with measurement data. Herein, we demonstrate how PINN could be used to improve WSS quantification in diseased arterial flows. Specifically, we assume we do not have any knowledge of the inlet boundary condition and incoming flow. We demonstrate that with very few measurement points collected even away from the vessel wall, we could use PINN to compute WSS with very high accuracy. We also show an interesting scenario where PINN does not correctly recover the inlet boundary condition but is still able to reconstruct WSS in the region of interest. We demonstrate examples in idealized stenosis and aneurysm models and discuss the implications of our model in transforming near-wall hemodynamics modeling.

More details in the arxiv paper:

https://arxiv.org/abs/2104.08249

Видео Near-wall Blood Flow Modeling with Physics-Informed Neural Network (PINN). канала Amirhossein Arzani