Enhanced Vascular Flow Simulations in Aortic Aneurysm via Physics-Informed Neural Networks and Deep Operator Networks

Oscar L. Cruz-González, Valérie Deplano, Badih Ghattas · 2025

Read Enhanced Vascular Flow Simulations in Aortic Aneurysm via Physics-Informed Neural Networks and Deep Operator Networks by Oscar L. Cruz-González, Valérie Deplano, Badih Ghattas (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-17402.txt.

Introduction

Oscar Cruz-González, Valérie Deplano, and Badih Ghattas speed up blood flow simulation in a 2025 paper. Aneurysms in the belly aorta need flow forecasts, but scans read flow poorly and classic simulation runs slow. Patient boundary data are also hard to get. The team tests learning based solvers that fold physics rules into neural networks, plus operator networks that learn solution maps whole.

Three model types faced a standard three dimensional aneurysm shape. Physics informed versions blend measured data with flow laws, which can cut the data each forecast needs. Faster, lighter forecasts could bring flow checks into routine planning for aneurysm repair. The comparison shows which learning route holds accuracy best for vessel flow.

Catalog record

CREATOROscar L. Cruz-González, Valérie Deplano, Badih Ghattas
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE101 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-17402

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Enhanced Vascular Flow Simulations in Aortic Aneurysm via Physics-Informed Neural Networks and Deep Operator Networks Oscar L. Cruz-González, Valérie Deplano, Badih Ghattas Abstract Due to the limited accuracy of 4D Magnetic Resonance Imaging (MRI) in identifying hemodynamics in cardiovascular diseases, the challenges in obtaining patient-specific flow boundary conditions, and the computationally demanding and time-consuming nature of Computational Fluid Dynamics (CFD) simulations, it is crucial to explore new data assimilation algorithms that offer possible alternatives to these limitations. In the present work, we study Physics-Informed Neural Networks (PINNs), Deep Operator Networks (DeepONets), and their Physics-Informed extensions (PI-DeepONets) in predicting vascular flow simulations in the context of a 3D Abdominal Aortic Aneurysm (AAA) idealized model. PINN is a technique that combines deep neural networks with the fundamental principles of physics, incorporating the physics laws, which are given as partial differential equations, directly into loss functions used during the training process. On the other hand, DeepONet is designed to learn nonlinear operators from data and is particularly useful in studying parametric partial differential equations (PDEs), e.g., families of PDEs with different source terms, boundary conditions, or initial conditions. Here, we adapt the approaches to address the particular use case of AAA by integrating the 3D Navier-Stokes equations (NSE) as the physical laws governing fluid dynamics. In addition, we follow best practices to enhance the capabilities of the models by effectively capturing the underlying physics of the problem under study. The advantages and limitations of each approach are highlighted through a series of relevant application cases. We validate our results by comparing them with CFD simulations for benchmark datasets, demonstrating good agreements and emphasizing those cases where improvements in computational efficiency are observed. The proposed methodology serves as a starting point for future research in the application of Deep Learning in cardiovascular disease modeling and offers a promising alternative for real-time simulation and monitoring of vascular flow in clinical settings. 1 Introduction Abdominal Aortic Aneurysms (AAA) are a critical health concern, characterized by the enlargement of the aorta diameter in the abdominal region. Accurate and efficient simulation of vascular flow within AAA is essential for predicting rupture risks and planning surgical interventions. Recent developments in magnetic resonance technology have led to the increased application of 4D flow MRI for analyzing hemodynamics. This imaging technique allows for the measurement of three-dimensional blood flow fields. Despite its benefits, the applicability to assess blood flow near vessel walls is challenged by limitations in spatial and temporal resolution, as well as the limited accuracy in analyzing slow-moving blood (see [ 1 ] ). On the other hand, computational methods such as Computational Fluid Dynamics (CFD), which rely on Navier-Stokes equations to simulate blood flow coupled with elastodynamic equations to model the arterial wall behavior have been widely adopted to model complex flow patterns in both large and small vessels with pathological changes. Using patient-specific inflow and outflow boundary conditions as well as patient-specific structure BCs, CFD simulations, taking into account fluid structure interaction (FSI), can deliver high-resolution insights into the elastohemodynamics analysis. Additionally, FSI enables the evaluation of pressure, shear stress and wall strains and stresses which are otherwise challenging to measure directly in living patients. Despite these benefits, the accuracy of these complex simulations is largely dependent on having access to detailed flow and structure BCs for all relevant vessels, which may be difficult to obtain for practical or ethical reasons. Moreover, CFD simulations accounting of such fluid structure interaction is time-intensive and require considerable computational resources, reducing their suitability for real-time clinical use [ 2 ] . So then, it is crucial to explore new data assimilation algorithms that offer possible alternatives to these limitations. Deep learning to achieve super-resolution and denoising of 4D flow MRI data, prediction of CFD simulations or combining approaches to build improved models are emerging areas of research and development. For instance, the authors in [ 3 ] introduced a framework that embeds the governing Navier–Stokes equations into a deep neural network to simultaneously achieve super-resolution and denoising without requiring high-resolution labels. Building on this concept, [ 4 ] developed a convolutional neural network approach that enforces physical constraints to recover high-fidelity flow fields from low-resolution inputs, further demonstrating the potential of label-free training in practical imaging scenarios. [ 5 ] proposed an implicit neural representation method for unsupervised super-resolution and denoising of 4D flow MRI data, offering a computationally efficient solution that inherently captures the underlying physics. Additionally in [ 6 ] , machine learning techniques, without relying on physics-informed methods, were used to develop DNNs that directly estimate 3D steady-state pressure and flow velocity distributions in a simplified model of the thoracic aorta. Similarly, [ 7 ] applied a dual-channel deep learning network along with cardiovascular hemodynamic point datasets in 3D to model the relationship between cardiovascular geometry and internal hemodynamics. Statistical analysis confirmed that the deep learning predictions matched conventional CFD results, while reducing computational time. [ 8 ] combined CFD simulations and deep learning to build a model that reduces noise and enhances the resolution of 4D Flow MRI velocities. Recent advances have also seen the results of [ 9 ] , who introduced a data-driven approach to optimizing blood flow in coronary artery bypass grafts using a novel reduced-order model that combines proper orthogonal decomposition (POD) with neural networks. This method is applied to steady 3D flows, focusing on patient-specific geometries and addressing the challenge of varying Reynolds numbers within physiological ranges. In [ 10 ] , the authors applied Physics-Informed Neural Networks (PINNs) to solve inverse hemodynamics problems in 3D, especially useful when boundary information is lacking and high-quality blood flow measurements are difficult to obtain. Focusing on the ascending aorta and taking advantage of sparse and noisy 2D measurements, the study demonstrated robust and relatively accurate parameter estimations when using the method with simulated data, while the velocity reconstruction accuracy shows dependence on the measurement quality and the flow pattern complexity. However, the study acknowledged its limitations, such as the assumption that the walls are rigid and the lack of validation with real clinical data. The reference solution is built using the finite element method (FEM). Additionally, in [ 11 ] further extensions to PINN were made to model incompressible flows in 2D for idealized geometries and time-dependent moving boundaries, a major challenge in current methodologies. By incorporating Dirichlet velocity constraints and refining the training points around the moving boundaries, the study improved the accuracy and applicability of no-slip conditions. This extended PINNs version demonstrated its effectiveness in solving unsteady flow problems and the potential for inverse problems, highlighting an area for future improvement. In [ 12 ] , an innovative physics-informed deep learning framework was introduced with the aim to solve steady-state incompressible flows over 2D irregular geometries. By employing point cloud neural networks and formulating loss functions based on governing equations and sparse observations, the framework surpassed the limitations of traditional PINNs, allowing solutions across multiple computational domains with significant geometric variations. The adaptability of the methodology to unknown geometries promises significant computational savings and broad applicability in modeling irregular cardiovascular geometries. In [ 13 ] , the authors presented a new computational framework called WS-PINNs, where WS refers to Warm-Start, to

Read the complete text