Enhancing Arterial Blood Flow Simulations through Physics-Informed Neural Networks
Shivam Bhargava, Nagaiah Chamakuri
TL;DR
This work tackles efficient arterial blood flow simulation by solving the incompressible Navier–Stokes equations with mesh-free Physics-Informed Neural Networks (PINNs). It introduces weighted XPINN (WXPINN) and weighted CPINN (WCPINN) to enable domain-decomposed, parallel training across subdomains, employing interface and, for CPINN, flux residuals to enforce physical continuity and conservation. The approach is demonstrated on rectangular and semi-circular arterial geometries, showing absence of backflow instabilities and improved convergence over standard PINNs, with systematic tuning of weights (\beta, \gamma, \delta) to balance governing equations, interfaces, and flux across subdomains. These results indicate a path toward high-fidelity, mesh-free hemodynamics simulations that can support clinical and biomedical engineering applications, by integrating domain-decomposed neural networks with physics-based training.
Abstract
This study introduces a computational approach leveraging Physics-Informed Neural Networks (PINNs) for the efficient computation of arterial blood flows, particularly focusing on solving the incompressible Navier-Stokes equations by using the domain decomposition technique. Unlike conventional computational fluid dynamics methods, PINNs offer advantages by eliminating the need for discretized meshes and enabling the direct solution of partial differential equations (PDEs). In this paper, we propose the weighted Extended Physics-Informed Neural Networks (WXPINNs) and weighted Conservative Physics-Informed Neural Networks (WCPINNs), tailored for detailed hemodynamic simulations based on generalized space-time domain decomposition techniques. The inclusion of multiple neural networks enhances the representation capacity of the weighted PINN methods. Furthermore, the weighted PINNs can be efficiently trained in parallel computing frameworks by employing separate neural networks for each sub-domain. We show that PINNs simulation results circumvent backflow instabilities, underscoring a notable advantage of employing PINNs over traditional numerical methods to solve such complex blood flow models. They naturally address such challenges within their formulations. The presented numerical results demonstrate that the proposed weighted PINNs outperform traditional PINNs settings, where sub-PINNs are applied to each subdomain separately. This study contributes to the integration of deep learning methodologies with fluid mechanics, paving the way for accurate and efficient high-fidelity simulations in biomedical applications, particularly in modeling arterial blood flow.