3–5 Jun 2026
Pisa
Europe/Rome timezone

Learning Vascular Hemodynamics: Physics-Informed GNNs Inference from Capillary Networks to Pulsatile Flow

4 Jun 2026, 11:15
15m
Aula E

Aula E

MS03 - Graph Neural Networks for Computational Physics MS03 - Graph Neural Networks for Computational Physics

Speaker

Paolo Botta (MOX, Department of Mathematics, Politecnico di Milano)

Description

Simulating microvascular blood flow in anatomically realistic networks remains a formidable computational task. The intrinsic multiscale structure of the vascular system, the geometric heterogeneity of capillary beds, and the nonlinear rheological behavior of blood in the microcirculation jointly lead to highly complex mathematical models whose direct numerical solution is often prohibitively expensive.

In this work, we propose a graph-based, physics-informed learning framework that exploits graph neural networks (GNNs) trained on synthetically generated microvascular graphs to approximate pressure and velocity fields with high efficiency. The approach integrates stochastic generative algorithms for vascular network construction with a physics-driven loss formulation enforcing mass conservation and rheological consistency. By embedding these domain-specific inductive biases into the learning process, the surrogate model preserves the governing physical structure while significantly reducing computational cost.

The resulting GNN architecture exhibits strong generalization capabilities across heterogeneous microvascular topologies and accurately reproduces full-order solutions under both linear and nonlinear rheological regimes. Substantial computational speedups are achieved without compromising predictive fidelity. Validation experiments conducted on anatomically reconstructed mouse cortical microvasculature further demonstrate the scalability, robustness, and reliability of the method in realistic settings.

As a natural extension, we address time-dependent hemodynamics. We introduce a complementary GNN architecture designed to reconstruct periodic flow regimes and to propagate pulsatile pressure and velocity dynamics across successive cardiac cycles on vascular graphs. The integration of steady and time-periodic formulations highlights the versatility of physics-informed graph learning for modeling complex vascular phenomena. Overall, the proposed framework provides a scalable and efficient paradigm for real-time simulation of both stationary and pulsatile microvascular blood flow, paving the way for advanced multiscale vascular modeling and biomedical applications.

Authors

Paolo Botta (MOX, Department of Mathematics, Politecnico di Milano) Dr Piermario Vitullo (MOX, Department of Mathematics, Politecnico di Milano)

Co-authors

Prof. Thomas Ventimiglia (Department of Biomedical Engineering, University of Illinois Chicago) Prof. Andreas Linninger (Department of Biomedical Engineering, University of Illinois Chicago) Prof. Lucas Müller (Laboratory of Mathematics for Biology and Medicine, Department of Mathematics, University of Trento) Prof. Paolo Zunino (MOX, Department of Mathematics, Politecnico di Milano)

Presentation materials

There are no materials yet.