Date of Award
Master of Science (MS)
Biomedical Engineering and Sciences
Venkat Keshav Chivukula
Ryan Tyler White
Abstract—Objective: We propose a new neural network architecture that accepts point clouds and outputs 3D velocity profiles for aneurysm geometries. Methods: We generated a synthetic aneurysm 3D flow dataset using CFD and used it to train our model architecture and compare it with other popular architectures like U-net and PointNet. We incorporate tools for improving model performance such as incorporating a distance function, a physics-informed loss to enforce the law of mass conservation, and the Huber loss to learn patterns across heterogeneous velocity components of multiple dimensions. Results: The tools implemented together with our architecture achieved the best performance on a novel synthetic aneurysm dataset, with an overall error of 7%. The 3D flow patterns of the prediction described by streamlines are in good agreement with the streamline plots of the ground truth CFD data. Our model predicts 3D flow patterns in less than a second for unseen geometries, multiple orders of magnitude faster than traditional CFD approaches. Significance: 3D flow patterns within blood vessels are important to physicians when assessing treatment of cardiovascular disease, which is a world-wide health concern. A neural network that can accurately predict blood flow in a vessel geometry in a short time would constitute the first step towards providing key information for clinicians to make better treatment decisions. Due to the intuitive and simple nature of the tools implemented, we have termed this best performing network FLIPINN: Fluid fLow Interpretable Physics Informed Neural Network.
Mattei Di Eugenio, Marcello Vittorio, "Development of a Physics-Informed Neural Network for Prediction of Blood Flow" (2023). Theses and Dissertations. 1379.
Available for download on Monday, December 16, 2024