Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
BP: Fachverband Biologische Physik
BP 32: Computational Biophysics II
BP 32.11: Vortrag
Freitag, 21. März 2025, 12:15–12:30, H46
Calibrating 1D-0D Coupled Blood Flow Models: the potential of Neural Network based Surrogates — •Benedikt Hoock1,2 and Tobias Köppl3 — 1Technische Universität München, School of Computation, Information and Technology — 2Support by Computing Facilities of Leibniz-Rechenzentrum München — 3Fraunhofer-Institut FOKUS, Berlin
Hydrodynamic models of the human arterial network can simulate the blood flow in parts or the whole body. The calculations can be simplified by solving the incompressible one-dimensional Navier-Stokes equations only for a set of larger vessels and coupling those at their outlets to a Windkessel model (1D-0D approach). Here, the right parametrization of the Windkessel parameters, i.e., the resistances and capacities, is crucial to obtaining realistic simulations. This can be done by calibrating the model parameters to match the model predictions with in-vivo blood pressure measurements. Since this requires many computationally expensive model evaluations, we test the potential of surrogates based on neural networks (NN). Once set up in an appropriate architecture, already ordinary fully connected NNs of moderate depth and width two can reproduce the simulations with high accuracy, advancing over e. g. the PINN approach due to their better trainability. We use these in an optimization algorithm to identify the target resistance and capacity with high precision in several test cases. Our efficient calibration scheme is an essential building block for an instantaneous visualization of the organ perfusion in a digital twin of a patient under different motion conditions on a digital treadmill.
Keywords: Bloodflow; Neural Networks; Windkessel Parameters; Optimization Techniques; Hydrodynamics