Surrogate modeling of flows in hemodynamics and aerodynamics using machine learning

Abstract

This thesis investigates the integration of traditional physics-based simulations and modern machine learning (ML) techniques for modeling fluid dynamics in both biomedical and engineering domains. The primary motivation stems from the computational limitations of high-fidelity simulations, particularly when exploring wide parametric spaces such as stenosis geometries in arteries. The study begins with a detailed Computational Fluid Dynamics (CFD) analysis of blood flow through idealized stenosed vessels. Stenosis geometry was parametrized using length (L), height (N), and shape exponent (n), and simulations were performed using both ellipsoid and superellipsoid profiles. The results demonstrated that stenosis height (N) has the most significant impact on hemodynamic parameters such as wall shear stress and pressure drop, with changes exceeding 300% in some configurations. However, the high computational cost associated with these simulations highlighted the need for surrogate models capable of rapid prediction. To bridge this gap, an intermediate ML study was conducted using NACA 4- and 5-digit airfoil data. This stage served as a controlled environment to understand the ML modeling pipeline, from dataset design and input encoding to hyperparameter tuning and generalization analysis. A compact, rule-based input encoding was developed from NACA designations and flow conditions, allowing neural networks to predict aerodynamic coefficients efficiently. The insights gained here, in areas such as dataset sensitivity, activation function choice, and model complexity, provided a solid foundation for applying ML to biomedical flows. Building on this groundwork, the final phase of the thesis applied ML models to predict hemodynamic quantities in stenosed vessels directly from geometric parameters. Various algorithms, including neural networks, support vector regression, and ensemble models, were trained on CFD-generated data. Neural networks showed the highest accuracy and strongest sensitivity to dataset size, making them particularly promising for biomedical prediction tasks. The resulting surrogate models significantly reduced computational requirements while maintaining high predictive fidelity, enabling scalable evaluations suitable for real-time clinical decision-making or large parametric studies. Overall, this thesis demonstrates a three-phase strategy: using CFD for physical insight, NACA-based ML for model development, and stenosis-based ML for biomedical application. The result is a flexible and robust modeling framework that leverages the strengths of both simulation and data-driven methods to accelerate research in fluid dynamics.