The accurate simulation of blood flow in the human vasculature is essential to understand the mechanical behavior of the cardiovascular system under both physiological and pathological conditions. Traditional numerical solvers, although highly accurate, become computationally expensive when applied to complex vascular networks or long transient regimes, thus limiting their use in large-scale or real-time analyses. This thesis introduces a data-driven framework based on Graph Neural Networks (GNNs) for the efficient and accurate prediction of pulsatile blood flow in synthetic vascular networks. The proposed model is composed of two complementary architectures: the Loader GNN, which reconstructs the system state at the onset of periodicity, and the Periodic GNN, which learns to propagate the hemodynamic variables over subsequent cardiac cycles. Their combination, referred to as the Full GNN, enables end-to-end prediction of the temporal evolution of pressure, flow rate, and vessel area throughout the vascular domain. The GNN framework is trained on high-fidelity data generated by a one-dimensional blood flow solver applied to a large collection of vascular geometries generated by suitable algorithms. Extensive testing demonstrates that the model can reproduce the full periodic dynamics with remarkable accuracy while reducing computational costs by several orders of magnitude compared to the full-order model (FOM). Moreover, the approach exhibits promising generalization capabilities across unseen vascular topologies. This work highlights the potential of graph-based learning techniques for hemodynamic simulations, paving the way toward real-time, scalable, and physiologically consistent modeling of blood flow in large and complex vascular networks.
La simulazione accurata del flusso sanguigno nel sistema vascolare umano è essenziale per comprendere il comportamento meccanico del sistema cardiovascolare in condizioni sia fisiologiche che patologiche. I solutori numerici tradizionali, sebbene altamente accurati, diventano computazionalmente costosi quando applicati a reti vascolari complesse o a regimi transitori di lunga durata, limitando così il loro utilizzo in analisi su larga scala o in tempo reale. Questa tesi introduce un framework data-driven che utilizza le Graph Neural Network (GNN) per la previsione efficiente e accurata del flusso sanguigno pulsatile in reti vascolari sintetiche. Il modello proposto è composto da due architetture complementari: la Loader GNN, che ricostruisce lo stato del sistema all'inizio della periodicità, e la Periodic GNN, che impara a propagare le variabili emodinamiche nei cicli cardiaci successivi. La loro combinazione, denominata Full GNN, consente la previsione end-to-end dell'evoluzione temporale della pressione, del flusso e dell'area dei vasi in tutto il dominio vascolare. Il framework GNN è addestrato su dati ad alta fedeltà generati da un risolutore di flusso sanguigno unidimensionale applicato ad una vasta famiglia di geometrie vascolari generate da opportuni algoritmi. Test approfonditi dimostrano che il modello è in grado di riprodurre l'intera dinamica periodica con notevole accuratezza, riducendo al contempo il costo computazionale di diversi ordini di grandezza rispetto al modello a ordine completo (FOM). Inoltre, l'approccio mostra promettenti capacità di generalizzazione su topologie vascolari non utilizzate nella fase di allenamento. Questo lavoro evidenzia il potenziale delle tecniche di apprendimento basate su grafi per le simulazioni emodinamiche, aprendo la strada alla modellizzazione in tempo reale, scalabile e fisiologicamente coerente del flusso sanguigno in reti vascolari grandi e complesse.
Learning pulsatile microvascular blood flow models using graph neural networks
BEHRENS, FRANCESCA
2024/2025
Abstract
The accurate simulation of blood flow in the human vasculature is essential to understand the mechanical behavior of the cardiovascular system under both physiological and pathological conditions. Traditional numerical solvers, although highly accurate, become computationally expensive when applied to complex vascular networks or long transient regimes, thus limiting their use in large-scale or real-time analyses. This thesis introduces a data-driven framework based on Graph Neural Networks (GNNs) for the efficient and accurate prediction of pulsatile blood flow in synthetic vascular networks. The proposed model is composed of two complementary architectures: the Loader GNN, which reconstructs the system state at the onset of periodicity, and the Periodic GNN, which learns to propagate the hemodynamic variables over subsequent cardiac cycles. Their combination, referred to as the Full GNN, enables end-to-end prediction of the temporal evolution of pressure, flow rate, and vessel area throughout the vascular domain. The GNN framework is trained on high-fidelity data generated by a one-dimensional blood flow solver applied to a large collection of vascular geometries generated by suitable algorithms. Extensive testing demonstrates that the model can reproduce the full periodic dynamics with remarkable accuracy while reducing computational costs by several orders of magnitude compared to the full-order model (FOM). Moreover, the approach exhibits promising generalization capabilities across unseen vascular topologies. This work highlights the potential of graph-based learning techniques for hemodynamic simulations, paving the way toward real-time, scalable, and physiologically consistent modeling of blood flow in large and complex vascular networks.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Behrens.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Testo della tesi
Dimensione
19.23 MB
Formato
Adobe PDF
|
19.23 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Behrens_Executive Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary
Dimensione
2.19 MB
Formato
Adobe PDF
|
2.19 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/247383