Aircraft in-flight icing is a serious meteorological threat that can lead to serious accidents. Anticipation of the effects of icing can be achieved by numerical simulation. This means of investigation uses a coupling between the air flow around the aircraft, often modeled by a RANS (Reynolds Averaged Navier-Stokes) approach and the solidification of atmospheric water droplets on the aircraft. Experimental observations show that the first moments of icing create surface roughness, often uncertain and irregular. Since roughness plays an important role in heat transfer, these uncertainties harm the quality of numerical predictions. Semiempirical roughness models are then used in the literature, but very few published works address the fine calibration of roughness for given atmospheric conditions. Published experimental accretions show the final ice shape, but the initial surface roughness is often unknown. Thus, the objective of the thesis is to suggest, develop and validate a methodology allowing the estimation of the roughness parameters to be entered into a RANS simulation by simply knowing the experimental ice shape. The methodology used consists of simulating the air flow with a compressible RANS solver with surface roughness. Subsequently, ice accretion is obtained with an icing solver based on the solidification of a runback liquid film, inspired by the shallow water icing model (SWIM). The droplets impingement relies on manufactured solutions specific to the wing geometry used, here a NACA0012 profile. The variation in roughness entered into the RANS solver ultimately produces a variation in ice accretion. More specifically, the project begins with a sensitivity study via Sobol indices and the creation of polynomial chaos metamodels to quantify the sensitivity of a 2D ice accretion to surface roughness parameters. The second step is the roughness calibration by Bayesian inversion to determine the roughness parameters leading to a prediction of heat transfers in line with the experimental observations. The implementation of a 3D ice accretion solver in SU2-CFD subsequently allows the methodology to be applied to 3D accretions. Finally, the method is extended to non-constant roughness distributions to refine ice accretion predictions and demonstrate the flexibility of the method. The calibrated roughness parameters obtained with the method yield ice accretions with less than 1.5 mm root mean square error compared to the experimental accretions, which represents 0.2% of the chord length. The ice shapes obtained also make it possible to show the interest of moving to a non-constant roughness distribution, since the results are greatly improved. In the four cases of icing with runback liquid film studied, the accretions obtained after calibrations are very close to the experimental shapes and sometimes more faithful than the numerical predictions published in the literature. The main outcome of this work is to make it possible to overcome the semi-empirical correlations usually used in the literature by estimating a tailor-made roughness for a specific test case. A future recommendation would be to calibrate more cases, in order to extract a specific roughness correlation for the models used and thus no longer have to systematically calibrate each case.
Keywords:
numerical simulation; aircraft icing; calibration; roughness; Bayesian inversion; SU2-CFD; polynomial chaos expansion metamodel; sensitivity