Data-Driven Surrogate Modeling for High-Fidelity Wall Shear Stress Assessment in Idealized Cerebral Aneurysms

Resumo

This article presents a comprehensive reduced-order modeling (ROM) framework for the rapid and accurate prediction of Wall Shear Stress (WSS) fields in idealized cerebral aneurysm geometries. Utilizing Proper Orthogonal Decomposition (POD) for efficient dimensionality reduction and a feedforward Neural Network (NN) as a surrogate model, the framework significantly reduces the computational expense of high-fidelity simulations. Key findings demonstrate that POD effectively compresses the solution space, capturing over 99\% of energy with a minimal number of modes, while the NN accurately predicts POD coefficients, achieving high R-squared values. This ROM pipeline enables WSS field predictions in milliseconds, offering substantial computational efficiency without compromising accuracy. The model successfully captures the geometric sensitivity of WSS patterns, including localized peak stresses at the aneurysm dome. Clinically, this framework holds promise for accelerating rupture risk assessment, optimizing treatment planning through virtual testing, and facilitating high-throughput morphological analysis. Current limitations include simplified geometric representations, static WSS estimates, and reliance on high-fidelity training data. Future work will focus on integrating patient-specific geometries, incorporating transient hemodynamics and fluid-structure interaction, and implementing uncertainty quantification for enhanced clinical applicability.