MeshGraphNet for Reduced-Order cardiovascular simulations

This example implements the one-dimensional Reduced-Order model based on MeshGraphNet presented in the paper Learning Reduced-Order Models for Cardiovascular Simulations with Graph Neural Networks (Pegolotti et al, 2023).

Three-dimensional simulations of the Navier-Stokes equations are the gold standard when it comes to modeling blood flow in arteries. However, these simulations are typically expensive, and a common way to alleviate the computational burden of evaluating physiological quantities of interest (e.g., pressure and flow rate) is using Reduced-Order models. For example, one-dimensional Reduced-Order models approximate the geometry of arteries as a composition of segments, the centerlines of the vessels, and the pressure and flow rate along the centerlines are found by solving special one-dimensional Partial Differential Equations (PDEs). These models are sometimes inaccurate due to their simplyfing assumptions.

We developed a one-dimensional Reduced-Order model able to mimic three-dimensional simulations accurately. The model is based on MeshGraphNet and trained on simulation of the 3D Navier-Stokes equations. As shown in the original reference, the model outperforms one-dimensional models in complex patient-specific cases featuring many junctions and/or pathological conditions.


Fig. 5 Comparison between the MeshGraphNet prediction and the ground truth for pressure and flow rate.

The dataset is composed of 310 simulations obtained on 8 different patient-specific models available in the Vascular Model Repository. Each simulation is stored as a .vtp file containing pressure and flow rate information at points located in the centerlines of the models and at different timesteps. The three-dimensional simulations were set up using SimVascular, an open-software software package for cardiovascular modeling and simulation, and run on 128 dual-socket AMD(R) EPYC 7742 cores of the San Diego Super Computing Center (SDSC) Expanse cluster. The simulations were obtained by varying inflow and outflow boundary conditions of each patient-specific model randomly.


Fig. 6 Patient-specific geometries contained in the dataset.

The base architecture is MeshGraphNet (see references for details). The node features of the graph neural network are:

  • pressure and flow rate at a particular timestep

  • cross-sectional area

  • tangent to the centerline

  • node type

  • cardiac cycle period in seconds

  • diastolic pressure

  • systolic pressure

  • RCR boundary condition parameters (only for outlet nodes)

  • loading variable (used to differentiate between an initial loading stage and the actual simulation phase)

The edge features are:

  • relative position of two nodes (in 3D)

  • distance between two nodes

  • edge type

The output of MeshGraphNet is the update in pressure and flow rate to get to the next timestep.

In order to deal with one-dimensional data, we made some modification to the original MeshGraphNet implementation. Most notably, we added special edges that connect boundary nodes to the interior one, to speed up the boundary condition information transfer.

Note: the default configuration for the architecture specified in config.yaml defines 64 as the dimension for hidden layers and outputs of encoder, processor and decoder. The results in the original paper were obtained by using 64 neurons in the hiddenl layers of each part of the network, and 16 neurons for the output layers of encoder and processor. This slight change in architecture does not influences the performance of the network dramatically.

To download the dataset (the vtp simulation files):


pip install gdown cd raw_dataset bash

After downloading the dataset, an intermediate step necessary to run MeshGraphNet is converting the simulation files into graphs compatible with DGL. This can be done with:


cd .. python

This will create a new graphs folder in raw_dataset. To train the model:



We currently support cpu and single-gpu training. The training parameters can be modified in config.yaml. An important parameter is training.geometries, which can take the values healthy, pathological, mixed. Here, healthy and pathological refer to the geometries used in Section 5.1 and 5.2 of the paper; mixed considers all geometries.

To perform inference on a given model:



The name of the model needs to be specified in config.yaml. Please refer to the list of graphs in raw_dataset\graphs for the possible graphs to use for inference.

The geometric data from the VMR is subject to license. See here for more information.

Previous RNN for transient 3D Gray Scott system
Next Molecular Dynamics using GNNs
© Copyright 2023, NVIDIA Modulus Team. Last updated on Jan 25, 2024.