Abstract
Simulating hemodynamics in elastic vessels is critical for cardiovascular disease diagnosis but remains computationally challenging for traditional mesh-based methods like the Finite Element Method (FEM). This paper presents a novel meshless solver based on Physics-Informed Neural Networks (PINNs) integrated with an Arbitrary Lagrangian-Eulerian (ALE) formulation. By decoupling the physics into a multi-network architecture and employing a sequential-alternative training strategy, the proposed ALE-PINN framework achieves a 10x speedup over FEM while maintaining high fidelity. The method demonstrates robustness in complex geometries, such as stenotic vessels with plaque, offering a flexible and efficient alternative for clinical blood flow modeling.
Introduction
The study of blood flow, or hemodynamics, is essential for understanding the progression of cardiovascular diseases like atherosclerosis. Accurate simulations must account for Fluid-Structure Interaction (FSI), where the pulsatile flow of blood deforms the elastic vessel walls. Traditionally, these problems are solved using FEM, which requires discretizing the domain into a mesh. However, in FSI problems, the mesh must be continuously updated to follow the moving boundaries, a process that is computationally expensive and prone to numerical instabilities, especially in complex or narrowed geometries.
To address these limitations, this research proposes a mesh-free approach using PINNs. By eliminating the need for a grid, this method simplifies the handling of moving boundaries and complex vessel shapes, providing a more scalable and flexible solution for hemodynamic modeling.
Methodology
ALE-PINN Framework
The core innovation is the integration of the Arbitrary Lagrangian-Eulerian (ALE) formulation into the PINN framework. The ALE method maps a fixed reference configuration $\Omega_0$ to a deformed current configuration $\Omega_t$ at any time $t$. This allows the neural network to learn the mapping of coordinates from a static domain to a moving one, effectively bypassing the need for mesh deformation.
Multi-Network Architecture
Rather than employing a single monolithic network, the solver utilizes three specialized neural networks to approximate different physical variables:
- Displacement Network ($N_d$): Predicts the deformation of the vessel wall and the internal fluid domain.
- Velocity Network ($N_u$): Computes the fluid velocity field.
- Pressure Network ($N_p$): Solves for the fluid pressure.
This decoupling allows each network to better capture the specific mathematical properties of its respective variable. The networks are trained by minimizing a composite loss function consisting of residuals from the Navier-Stokes equations, elasticity equations, and relevant boundary conditions.
Optimization Strategies
To ensure stability and convergence, the framework incorporates several advanced techniques:
- Hybrid Activations: A combination of ReLU and Sigmoid functions is used. Sigmoid provides the smoothness necessary for high-order derivatives in physics losses, while ReLU accelerates convergence.
- Sequential-Alternative Training: A specialized training schedule optimizes subsets of the loss function in sequence. This "freezing" of certain physical components during initial training phases prevents the optimization instabilities common in complex FSI-PINN models.
Results and Discussion
The ALE-PINN solver was evaluated against high-resolution FEM benchmarks across various scenarios, including healthy cylindrical vessels and vessels with varying degrees of plaque-induced stenosis.
Performance and Accuracy
The solver achieved a relative error for velocity magnitude of approximately 1%, matching the accuracy of dense-grid FEM simulations. Crucially, the PINN approach demonstrated a significant computational advantage, completing a typical simulation in 863 seconds compared to over 9,155 seconds for FEM—a greater than 10-fold speedup. Furthermore, the method showed excellent scalability, with training time nearly halving when distributed across 4 GPUs.
Clinical Implications
The meshless nature of the solver allows for the easy incorporation of complex anatomical features. Simulations of stenotic vessels successfully captured the "stenosis effect," characterized by increased wall stress and reduced blood flux in narrowed regions. This capability is vital for assessing the severity of arterial plaques and planning interventions.
Conclusion
The ALE-PINN framework represents a significant step forward in meshless hemodynamic simulation. By combining the flexibility of PINNs with the rigorous ALE formulation, the solver overcomes the traditional "meshing headaches" of FSI problems. Future work will focus on incorporating non-linear material models and extending the framework to 3D patient-specific geometries for inverse design applications, such as stent optimization.
Summary updated