Investigation of Aircraft Icing Based on Numerical Modeling and Machine Learning Methods

The goal of this work is to investigate the aircraft icing severity by using numerical modeling and machine learning methods. Since the ice buildup on the wing’s leading edge may alter the original aerodynamic configuration and degrade the aerodynamic performances, the ice shape is predicted by the numerical simulation approach. Numerical simulations generally require computationally expensive and/or cumbersome treatments to calculate the ice displacement and accretion along the wing, such as remeshing or other sophisticated techniques. Based on that fact, a prediction model based on machine learning methods is built to extract general yet critical ice features such as the ice global coverage or maximum ice thickness. The first stage of the work is comprised of two chapters, where the focus is modeling of ice accretion over aircraft wings. An approach based on the Eulerian two-phase flow theory to numerically simulate ice accretions on an aircraft wing is developed and implemented in OpenFOAM to contribute the open-source community. The airflow field is obtained by solving the Navier-Stokes equations. Turbulence can be modeled or resolved to improve the prediction accuracy, especially in evaluating the effect of icing on the aerodynamics. A permeable wall condition is proposed to simulate the droplets impingement properly. The droplets collection efficiency is calculated according to the droplets velocity and volume fraction. The ice accretion thermodynamic model is built based on the improved Messinger model and the wall function is modified to account for roughness effect in calculating the convective heat transfer coefficient. The ice shape is reconstructed based on the assumption that ice grows in the direction normal to the aircraft surface. The mesh morphing model adjusts the wing’s shape every time step based on the amount of ice accreted. Therefore, the air flow field is updating during the simulation. Ice accretion on a NACA0012 airfoil and an ONERA M6 wing model under different conditions have been simulated to validate the solver’s performance and investigate the effect of the accreted ice on the aerodynamic performance. Then, a density based compressible solver is proposed to improve the scaling performance for time accurate simulations. The aim of developing this solver is to twofold: (i) to improve the scaling performance of rhoCentralFoam, especially in large-scale simulations; and (ii) to improve time-accuracy and overall time-to-solution using high-order Runge-Kutta scheme. The parallel scalability of the solver is demonstrated through numerical experiments conducted on a Cray XC40 parallel system Theta, at Argonne National Laboratory. The proposed solver could be easily incorporated into the icing solver when conducting large eddy simulations. The second stage of the thesis introduces a data-driven statistical model for aircraft icing severity evaluation. The complex process of ice accretion makes machine learning-based methods an attractive alternative to experiments and traditional numerical simulation-based approaches. We adapt the multiple conventional and ensemble machine learning models to six atmospheric and flight conditions - liquid water content, median volumetric diameter, exposure time, static temperature, angle of attack and flight speed. The prediction models are demonstrated in three cases: maximum ice thickness prediction, icing area prediction and icing severity level evaluation. Multiple performance measures are employed and the results show that proposed data-driven model has a satisfactory capability to evaluate aircraft icing severity. The effects of two selected flight conditions on aircraft icing are further studied.