Leveraging Graph Neural Networks for Efficient Reduced-Order Blood Flow Simulations
DescriptionRecently, simulations of blood flow have shown great promise in revolutionizing cardiovascular disease research and treatment. Reduced-order models, specifically zero- and one-dimensional ones, can approximate blood dynamics more efficiently than detailed three-dimensional simulations. These models prove helpful when computational resources or time are constrained or in scenarios requiring numerous queries, such as uncertainty quantification. However, their accuracy can falter in complex geometries featuring many junctions or pathological conditions like stenoses or aneurysms. Data-driven reduced-order models utilize previous simulation data to address the limitations of conventional physics-based models. Our presentation delves into a one-dimensional reduced-order model utilizing MeshGraphNet, a graph neural network architecture designed for simulations on unstructured grids. The graph neural network is trained on various three-dimensional simulation data restricted to the nodes on the geometry centerline. Thanks to the versatility of graph neural networks, a single trained architecture can effectively approximate blood dynamics in a range of topologies and geometries. Our results show that, given sufficient training data, our algorithm outperforms physics-based one-dimensional models in accuracy while still offering dramatic speed enhancements compared to three-dimensional simulations. In this presentation, we also cover the limitations of the current methodology and future perspectives.
TimeTuesday, June 2711:30 - 12:00 CEST