Learning Vascular Hemodynamics: Physics-Informed GNN Inference From Capillary Networks to Pulsatile Flow
Please login to view abstract download link
Simulating blood flow in realistic cerebral microvascular networks is a cornerstone for the development of high-fidelity biomedical digital twins, yet it remains computationally prohibitive due to multiscale effects, complex vascular topology, and nonlinear blood rheology. We present a graph-based, physics-informed learning framework that leverages graph neural networks (GNNs) trained on synthetic microvascular graphs to efficiently approximate pressure and velocity fields. The method combines generative algorithms for vascular network construction with a physics-based loss function that enforces mass conservation and rheological consistency, thereby incorporating domain-specific inductive biases into the surrogate model. The resulting GNN architecture generalizes robustly across different microvascular topologies and accurately reproduces full-order solutions for both linear and nonlinear rheological models, achieving significant increases in computational speed. Validation on mouse cortical microvasculature highlights the scalability and reliability of the approach on anatomically realistic domains. To move toward fully dynamic digital twins, we further extend the framework to time-dependent hemodynamics. A complementary GNN architecture is introduced to learn periodic flow regimes and propagate pulsatile pressure and flow waves across cardiac cycles, trained against high-fidelity transient simulations based on one-dimensional blood flow models. Overall, this work demonstrates the potential of physics-informed graph learning as a scalable and efficient core technology for constructing cerebral microvascular digital twins of the mouse brain, opening new avenues for multiscale modeling, in silico experimentation, and translational neuroscience.
