A dual-branch physics-informed neural network for fluid-structure interaction analysis of aortic valve

Aortic valve stenosis is a prevalent cardiac valvular disease, and accurate estimation of the coupled dynamics of the valve and blood flow is essential for early diagnosis. Currently, medical imaging provides insights but is limited by the imaging constraints and the inability to quantify stresses, while fluid-structure interaction (FSI) models face slow convergence and uncertain boundary conditions. Recently, physics-informed neural networks (PINNs) have shown promise in hemodynamic problems but remain limited by low accuracy and loss imbalance in complex valve FSI. To address these limitations, this study proposes a physics-informed neural network for fluid-structure interaction (FSIPINN) for the kinetic parameters of the valve and associated blood flow. FSIPINN integrates the governing equations of fluid and structure domains into a dual-branch architecture, with an adaptive weighting strategy to balance domain disparities and data-physics contributions. The fluid equations are reformulated into momentum conservation and constitutive relations to directly estimate shear stress, reduce derivative order, and accelerate convergence. FSIPINN was evaluated on aortic valve FSI simulations with varying stenosis severities. The relative L2 errors were 0.0350 ± 0.0195 for blood flow velocity and 0.363 ± 0.044 for the corresponding shear stress. Similarly, the relative L2 errors were 0.0037 ± 0.0019 for valve displacement and 0.268 ± 0.076 for the corresponding shear stress. Among the various training strategies considered, simultaneous training achieves the best performance. These results demonstrate that FSIPINN can provide relatively accurate spatiotemporal estimations of valve and blood flow parameters, indicating its potential as a promising tool for FSI analysis.