The aim of this study is to further investigate the accuracy and the reliability of the actuator line model (ALM) predictions for turbulent separated wakes. Large eddy simulations (LES) of the flow around a NACA0009 airfoil are performed mimicking the geometry with the immersed boundary method. Results are validated against experiments and used to assess the accuracy of the ALM predictions for the same airfoil, with different values of the spreading parameter and of the reference velocity and for two values of the angle of attack. It is found that the ALM setup recently derived from linearized inviscid analysis leads to accurate results for the lower angle of attack, while at the higher one for which a significant separation of the boundary layer occurs, the ALM requires a different set of model parameters. This calls for a systematic investigation of the sensitivity to the ALM parameters for separated flows, which is carried out herein through a stochastic approach allowing continuous response surfaces to be obtained in the parameter space. The ALM parameters are calibrated against the results obtained with the immersed boundaries. With the calibrated model parameters, the ALM gives good predictions of the velocity and turbulent kinetic energy in the far wake. Finally, the proposed model parameters are used to predict the flow past a different geometry, a flat plate, at high angle of attack.The accuracy of the prediction of the far wake is again good, showing the robustness of the identified setup. KEYWORDS actuator line model, generalized Polynomial Chaos expansion, stochastic sensitivity analysis, turbulent wake 1In the actuator line model (ALM), the force is distributed in a small region centered into the line representing the blade position. The spreading kernel is typically a Gaussian function of the distance from the actuator line position. The standard deviation of the Gaussian kernel, , determines the width of the spreading. The ALM is known to represent more accurately the wake flow features than similar models, like the actuator disk or the rotating actuator disk, which have lower computational cost. [3][4][5] However, the flow obtained with the ALM strongly depends on the spreading parameter ( ). There are no definitive guidelines to set this input parameter. For numerical stability, it is suggested that the spreading parameter be at least twice the grid spacing, in order to avoid singularities due to a pointwise force. 6 This was confirmed by Martínez et al. 3 who investigated the sensitivity to the spreading to grid-spacing ratio as well as to the grid resolution for a fixed spreading value. Shives and Crawford suggested 7 that the spreading parameter should be related to the local chord length ( ∕c ≃ 1∕8-1∕4) and thus vary along the radial direction for a common wind turbine blade geometry. An analogous approach was also pursued by Jha et al. 8 and Jha and Schmitz 9 defining an equivalent elliptic blade planform.Recently, Martìnez et al 10 carried out analytical computations using the linearized Eu...