Reduced-Order Surrogate Modeling of non-Newtonian blood stream inside Cerebral Aneurysm via hybrid POD-Transformer Networks

Accurate and efficient prediction of hemodynamic parameters is essential for assessing the rupture risk of cerebral aneurysms, particularly in the middle cerebral artery (MCA). In this study, we propose a reduced-order modeling framework that combines Proper Orthogonal Decomposition (POD) with a Transformer-based deep learning model to reconstruct and forecast key hemodynamic indicators—velocity components, wall shear stress (WSS), oscillatory shear index (OSI), and pressure (P)—across the pulsatile cardiac cycle. Patient-specific flow simulations were performed under age-dependent inlet conditions for two cohorts (ages 20–39 and 50–59 years) to evaluate the model's performance under physiologically realistic scenarios. POD was used to extract dominant flow features and reduce the dimensionality of the data, with energy analysis showing that fewer than 10 modes captured over 95% of the energy for velocity and pressure, while OSI required more modes due to its complex oscillatory behavior. The Transformer model, trained on the temporal evolution of the POD coefficients, accurately reproduced the flow field dynamics, achieving low reconstruction errors on both training and unseen test data. Comparative analysis revealed that the older age group exhibited more stable OSI behavior, resulting in improved predictive accuracy. In contrast, the younger group, characterized by more dynamic and energetic flow, required more modes and exhibited higher prediction error—particularly in OSI. Overall, the proposed POD + Transformer framework enables fast and reliable prediction of critical hemodynamic factors with a significant reduction in computational cost. This approach holds promise for large-scale aneurysm studies and real-time clinical applications in personalized risk assessment and treatment planning.