Search

This article discusses an application of Machine Learning (ML) tools to improve the prediction accuracy of Computational Fluid Dynamics (CFD) for external aerodynamic workflows. The Reynolds Averaged Navier-Stokes (RANS) approach to CFD has proved to be one of the most popular simulation methodologies due to its quick turnaround times and acceptable level of accuracy for most applications. However, in many cases the accuracy for the RANS models can prove to be suboptimal that can be significantly improved with model closure coefficient tuning. During the original turbulence model creation, these closure coefficients were chosen by somewhat ad hoc methods using simple canonical flows that do not transfer well to flows involving more complex objects, like the automotive bodies used in this work. This work presents a novel method of applying ML tools to CFD to optimize the turbulence closure coefficients by using model explainability tools such as Shapley Values, Shapley Additive exPlanations (SHAP), and ML surrogate models. The 25-degree slant Ahmed body model was used to obtain sampling data to tune closure coefficient in the Menter Shear Stress Transport (SST) turbulence model implemented in the open source CFD code, OpenFOAM v2012. Shapley additive values were then calculated using the samples which showed that β∗ has the strongest influence over the model predictions of lift and drag. ML surrogate models were then applied alongside SHAP providing a better overall sampling efficiency with Shapley additive values and more complete explanations of the model. The SHAP explanations showed that β∗ had the most influence on the force predictions followed by σω2, while σω1, σk1, and σk2 were shown to have little impact. The surrogate model was then used along with its explanations to provide optimized coefficients that reduced the error in the drag and lift predictions to -3.67% and -2.49% respectively, from -9.67% and -75.8%.

In spite of its shortcomings, faster turnaround time and cost-effectiveness make the Reynolds-averaged Navier–Stokes modeling approach still a popular and widely used methodology in many industrial applications, including the automotive industries in general, but the motorsports sector in particular. Existing literature suggests that all Reynolds-averaged Navier–Stokes models generally fail to predict pressure and velocity flow fields with a reasonable accuracy, especially in the vehicle wake region. Recent numerical works suggest that, when using two-equation eddy viscosity turbulence models, improved correlation between the experiment and computational fluid dynamics is not achievable through only mesh refinements or adding additional corrective terms in the turbulence transport equations, and additional efforts are necessary for better predictions. In this backdrop, the prediction improvement strategy adopted in this paper is based on the realization that the turbulence model closure coefficients are normally specified as constants and their values are determined from either a single observation or a simple functional form is assumed and that these coefficients are constructed and/or constrained to behave correctly in extremely limiting circumstances. Subsequently, this study investigated the influence of a few selected turbulence model closure coefficients on the veracity of computational fluid dynamics predictions by analyzing simulations run with turbulence model closure coefficient values that were different from the commonly used default ones. This was done by first investigating the individual effect of each model parameters on the prediction veracity, and then a combination of model closure coefficient values was formulated in order to obtain a prediction with the best experimental correlation. This procedure was applied to four different test objects which include NACA 4412 airfoil at 12° angle of attack, the 25 and 35° slant angle Ahmed body, and a full-scale sedan type passenger vehicle. The shear-stress transport [Formula: see text] was chosen as the turbulence model because of its popularity in automotive applications. It was observed that, irrespective of the test model, the computational fluid dynamics predictions are somewhat independent of some closure coefficients, while the prediction showed strong dependencies on certain model constant values, especially those in the dissipation equation. The present study demonstrated that, by using a modified set of closure coefficient values, very well correlated computational fluid dynamics predictions were possible. The findings from this study simply reiterate a 20-year-old conclusion by Stephen Pope that the closure coefficients for a model are not static and universal but instead are dynamic and must be calibrated for specific flow situations.

This paper presents a study on the flow dynamics involving vehicle interactions. In order to do so, this study first explores aerodynamic prediction capabilities of popular turbulence models used in computational fluid dynamics simulations involving tandem objects and thus, ultimately presents a framework for CFD simulations of ground vehicle platooning using a realistic vehicle model, DrivAer. Considering the availability of experimental data, the simulation methodology is first developed using a tandem arrangement of surface-mounted cubes which requires an understanding on the role of turbulence models and the impacts of the associated turbulence model closure coefficients on the prediction veracity. It was observed that the prediction accuracy of the SST k−ω turbulence model can be significantly improved through the use of a combination of modified values for the closure coefficients. Additionally, the initial validation studies reveal the inability of the Unsteady Reynolds-Averaged Navier-Stokes (URANS) approach to resolve the far wake, and its frailty in simulating tandem body interactions. The Improved Delayed Detached Eddy Simulations (IDDES) approach can resolve the wakes with a reasonable accuracy. The validated simulation methodology is then applied to the fastback DrivAer model at different longitudinal spacing. The results show that, as the longitudinal spacing is reduced, the trailing car’s drag is increased while the leading car’s drag is decreased which supports prior explanations of vortex impingement as the reason for drag changes. Additionally, unlike the case of platooning involving Ahmed bodies, the trailing model drag does not return to an isolated state value at a two car-length separation. However, the impact of the resolution of the far wake of a detailed DrivAer model, and its implication on the CFD characterization of vehicle interaction aerodynamics need further investigations.

In today&rsquo, s road vehicle design processes, Computational Fluid Dynamics (CFD) has emerged as one of the major investigative tools for aerodynamics analyses. The age-old CFD methodology based on the Reynolds Averaged Navier&ndash, Stokes (RANS) approach is still considered as the most popular turbulence modeling approach in automotive industries due to its acceptable accuracy and affordable computational cost for predicting flows involving complex geometries. This popular use of RANS still persists in spite of the well-known fact that, for automotive flows, RANS turbulence models often fail to characterize the associated flow-field properly. It is even true that more often, the RANS approach fails to predict correct integral aerodynamic quantities like lift, drag, or moment coefficients, and as such, they are used to assess the relative magnitude and direction of a trend. Moreover, even for such purposes, notable disagreements generally exist between results predicted by different RANS models. Thanks to fast advances in computer technology, increasing popularity has been seen in the use of the hybrid Detached Eddy Simulation (DES), which blends the RANS approach with Large Eddy Simulation (LES). The DES methodology demonstrated a high potential of being more accurate and informative than the RANS approaches. Whilst evaluations of RANS and DES models on various applications are abundant in the literature, such evaluations on full-car models are relatively fewer. In this study, four RANS models that are widely used in engineering applications, i.e., the realizable k - &epsilon, two-layer, Abe&ndash, Kondoh&ndash, Nagano (AKN) k - &epsilon, low-Reynolds, SST k - &omega, and V2F are evaluated on a full-scale passenger vehicle with two different front-end configurations. In addition, both cases are run with two DES models to assess the differences between the flow predictions obtained using RANS and DES.