Integrated CFD-based reduced-order modeling for ice accretion and aerodynamic performance predictions

Abstract

Aircraft icing represents one of the most critical hazards in aviation, posing a significant threat to flight safety and operational efficiency. The accretion of ice on aerodynamic surfaces, such as wings, tails, and control surfaces, fundamentally alters the geometry of the aircraft, leading to severe degradation in aerodynamic performance. Specifically, ice accumulation increases surface roughness and modifies the shape of the airfoil, resulting in a reduction in the maximum lift coefficient, a substantial increase in drag, and adverse changes in the characteristics of the pitching moment. These effects can drastically reduce the flight envelope, impair stability and control, and in extreme cases, lead to catastrophic flow separation and loss of control. Consequently, flight safety is considered a fundamental priority in aircraft design, certification, and operational practice. Regulatory authorities, including the Federal Aviation Administration (FAA) and the European Union Aviation Safety Agency (EASA), mandate rigorous certification requirements to ensure that aircraft can operate safely within defined icing envelopes. To comply with these regulations, aircraft designers must demonstrate the aircraft's performance under various icing conditions. This involves a combination of flight tests, icing wind tunnel campaigns, and numerical simulations. Although flight tests and wind tunnel experiments provide the most realistic data, they are prohibitively expensive, time-consuming, and subject to safety risks or facility limitations. As a result, numerical simulation tools have become indispensable in design, certification, and operational processes. High-fidelity Computational Fluid Dynamics (CFD) simulations are widely recognized as reliable tools for predicting ice accretion and its aerodynamic consequences. However, the simulation of ice accretion is a complex, multi-physics problem involving aerodynamics, droplet trajectories, thermodynamic heat balance, and phase changes. Solving these coupled equations for a wide range of flight conditions requires significant computational resources and time. This high computational cost makes high-fidelity CFD impractical for rapid analysis, extensive parametric studies, or integration into real-time engineering simulations. Addressing these challenges, this thesis develops an integrated Reduced-Order Modeling (ROM) framework designed for the rapid and accurate prediction of in-flight ice accretion and the associated aerodynamic performance. The primary objective is to bridge the gap between the high accuracy of CFD and the computational efficiency required for rapid engineering analysis. The proposed framework combines physics-based CFD simulations with advanced data-driven model reduction techniques in a modular and extensible structure. The methodology is built upon four tightly coupled components: (1) a baseline aerodynamic prediction model for clean configurations, (2) a supervised ice-type classification model, (3) dedicated reduced-order models for ice shape prediction, and (4) an aerodynamic degradation model for estimating performance losses due to ice accretion. The foundation of the proposed framework is a robust database generated using high-fidelity CFD simulations. The FENSAP-ICE solver is utilized to simulate the complex physics of ice accretion. The simulation process follows a multi-shot icing strategy to capture the time-dependent nature of ice growth. In this approach, the flow field is solved, droplet impingement is calculated, ice grows on the surface, and the computational mesh is updated (remeshed) to accommodate the new geometry. This cycle is repeated to simulate realistic ice buildup over time. Particular attention is given to modeling surface roughness and thermodynamic energy balance, which are critical to determining the type of ice formed—ranging from streamlined rime ice at low temperatures to complex, horn-shaped glaze ice at temperatures near the freezing point. These high-fidelity simulations ensure that the training data for the ROM capture the essential physics of the problem. A central contribution of this work is the development of a comprehensive ROM strategy based on Proper Orthogonal Decomposition (POD) combined with Kriging interpolation. POD is used to extract the dominant energy modes from the snapshot matrices of flow field variables and ice geometries. For aerodynamic prediction, surface pressure coefficients and wall shear stress components are decomposed into their modal contributions. This allows the complex flow field to be represented by a small number of modes and coefficients. However, applying POD to ice accretion presents specific challenges: the mesh topology changes as ice grows, resulting in varying grid sizes and node distributions for each sample. To overcome this, a novel redistribution methodology is introduced. Ice geometries obtained from variable CFD meshes are mapped onto a consistent, node-aligned reference grid using parameterization. This ensures that the topological structure of the data remains consistent, enabling effective modal decomposition under a wide range of icing conditions. The thesis demonstrates that ice accretion behavior is highly nonlinear and strongly dependent on the ice type (rime vs. glaze). Glaze ice, characterized by the formation of "horns" due to run-back water, exhibits significantly different geometric features compared to the smoother rime ice. Standard global ROM approaches often fail to capture these distinct topological differences accurately. To address this, a supervised machine learning classifier is integrated into the framework. This classifier uses input parameters such as ambient temperature, velocity, and liquid water content to predict the resulting ice type (rime, glaze, or mixed) before the ROM step. Based on this classification, the framework selects a specialized local ROM trained specifically for that ice regime. To ensure robustness of the model and adequate coverage of the design space, a sequential sampling strategy is employed. Initially, the parameter space is populated using Latin Hypercube Sampling to ensure a uniform distribution of sample points. Following the construction of the initial models, a Leave-One-Out Cross-Validation (LOOCV) analysis is performed to identify regions of high prediction error, which are typically associated with nonlinear transition boundaries between rime ice and glaze ice. Based on this error assessment, additional sample points are incorporated into these critical regions to enrich the dataset. This targeted approach ensures that the model achieves high accuracy throughout the entire envelope while maintaining computational efficiency in the data generation phase. The proposed framework is first developed and validated using a two-dimensional NACA 0012 airfoil. This case serves as a controlled environment for method verification. The results demonstrate that the ROM accurately reproduces ice shapes, aerodynamic coefficients, and icing-induced performance degradation over a broad icing envelope. The model demonstrates particular capability in capturing the aerodynamic characteristics in the pre-stall regime. Demonstrating the scalability of the method, the methodology is subsequently applied to the three-dimensional NASA Common Research Model (CRM), a representative transport aircraft configuration. The transition from two-dimensional to three-dimensional introduces significant complexity, including spanwise flow effects and swept-wing icing characteristics. High-fidelity CFD datasets are generated for the CRM wing and reduced-order models are constructed accordingly. The three-dimension results confirm that the framework effectively captures the spanwise variation of ice shapes and the resulting global aerodynamic degradation. A detailed computational efficiency analysis reveals the transformative potential of the developed framework. Although the offline phase involves computationally intensive CFD simulations and model training, this is a one-time cost. Once trained, the online prediction phase is extremely fast. The ROM provides full ice shape and aerodynamic performance predictions in less than one second per case on a standard workstation, representing a speed-up of several orders of magnitude compared to the hours required for equivalent full-order CFD simulations. In conclusion, this thesis presents a validated, extensible, and highly efficient CFD-based ROM framework. The framework establishes a strong foundation for future extensions, including unsteady icing simulations and integration into flight simulators for pilot training, thereby directly contributing to the enhancement of aviation safety.