NablaFlow tests and expands its abilities

Introduction

The AeroCloud development team is always looking ahead to any innovation that can enhance their tool. In particular, given the significant progress that CFD has made over the past decades, the implementation of advanced simulations, the transient simulations, is now more and more common. Transient simulations are CFD tools capable of providing more detailed and accurate results, compared to the standard steady-state simulations. Their main limitation lies in the computational cost: unlike steady-state cases, they often require simulation times that can be tens of times higher.

Detached Eddy Simulation (DES) techniques have been developed in recent years as hybrid approaches, combining elements of RANS (Reynolds-Averaged Navier–Stokes) and LES (Large Eddy Simulation). The goal is to merge the reduced computational cost of RANS with the high level of flow field resolution provided by LES.

The Project

The AeroCloud team aimed to investigate in detail the potential implementation of these technologies in their services through a Master’s thesis project in Mechanical Engineering, in collaboration with the renowned University of Bologna.

To this end, custom numerical schemes and a dedicated OpenFOAM solver were developed, capable not only of performing such simulations but also of significantly reducing computational costs by lowering run times.

The project was based on the Windsor body geometry, a well-known benchmark in the literature thanks to the experimental results obtained at the Loughborough Large Wind Tunnel, as well as the CFD results available on the AutoCFD portal, where research institutes and universities worldwide have contributed with transient simulations.

These events aimed to test different CFD techniques, generally with much stricter accuracy criteria compared to current industrial standards.

Challenges

The mesh used is a very coarse one, with around 5 million cells, corresponding to the basic quality level of AeroCloud. The study should therefore be considered a preliminary analysis; an exact match is not expected, but at least an initial correlation is anticipated.The project was structured into 2 simulation Sets, where each Set consisted of cases with progressively larger time-step sizes.

This approach aimed to reduce the total number of time intervals in the simulation and, consequently, the overall computational cost. One of the most challenging tasks was to identify the optimal time-step value, ensuring that the reduction in computational cost was still associated with accurate and reliable results.

Results

For the comparison were analyzed: the drag coefficient, the normalized mean velocity field acquired on three planes and three pressure measurement groups located on the body:

Regarding the drag coefficient, as can be seen in the figures, the results for Sets 1 and 2 generally show an underestimation of drag compared to the experimental reference:

In general, this underestimation tends to increase with higher temporal discretization, as normally expected. However, it can be observed that simulations characterized by the same ΔT value tend to produce almost identical results in terms of average drag and its variation.

A more interesting comparison can be made with the results obtained from the AutoCFD seminars:

Here, it becomes evident how complex and intricate the world of CFD simulations truly is: research institutions from all over the world have produced a very wide range of results: some simulations appear to match the experimental reference quite accurately, while others tend to significantly underestimate or overestimate the results. Focusing on these simulations in more detail, isolating only the transient results and grouping them by turbulence modeling approach, it can be observed that the RANS results (magenta) generally deviate from the experimental reference and are less accurate, the LES models (orange) are instead more precise and more closely clustered around the experimental value.

As shown in light blue, the hybrid DES techniques are effectively intermediate between the two behaviors, spreading across a wide range of results.

By analyzing the specific RANS model used within the DES approach, a reference term for AeroCloud’s simulation results is obtained: DES SST:

The fitting of the AeroCloud simulations particularly for the first simulation cases is remarkably good, especially when considering the low number of cells used.

Regarding the pressure measurements for Groups 1 and 2, all simulation cases produced similar results. The results associated with Group 1 are reported here:

All the simulations of Set 1 and 2 produced the same pressure coefficient distribution. This means that if the result is obtained independently of the time discretization, then variations in the time step (ΔT) do not have an effect in the region upstream or around the body immersed in the flow. As can be seen, there is a mismatch with some of the experimental references; however, the same discrepancy is also found in the reference results obtained from the AutoCFD seminars. This is probably due to intrinsic differences between the virtual environment and the experimental setup in the wind tunnel.

In the rear region of the Windsor body, there is a tendency to overestimate the pressure coefficient. This is consistent with the computational domain used in AeroCloud, which is larger than both the wind tunnel domain and the AutoCFD simulations, where the narrower region tend to accelerate more the flow, providing lower pressure values according to Bernoulli’s principle.

For the planar acquisition comparison, the results corresponding to one of the three planes the horizontal one are reported:

When compared with the experimental reference, it can be observed that the transient simulations, shown for increasing temporal discretization, reproduce fairly well the flow field downstream of the Windsor body, with a wake that appears consistent in both shape and intensity. Given the low number of cells, it is inevitable that the comparison with the reference may not be accurate; this is expected in a preliminary study. As can be seen, the recirculation bubble, characterized by the light blue coloring, is well represented in both intensity and shape; however, it appears slightly more detached from the body, i.e., positioned farther to the right.

A similar behavior can also be observed for the simulations obtained on the AutoCFD platform, though with a wider spread: the simulations more accurately reproduce the bubble position, but sometimes the intensity (areas of deep blue) or the shape is not fully captured:

This once again confirms that the transient DES technique can achieve a higher level of detail compared to steady-state simulations, although it is not exempt from producing particular or anomalous results.

So, are transient simulations with 5 million cells possible?

Unfortunately not, as demonstrated by the comparison with Group 3, which represents the rear face of the Windsor body:

Here, no clear agreement with the experimental reference is observed. For the same color scale, the results produced on AeroCloud show a strong underestimation of the pressure field, consistent with the overall underestimation observed in the drag coefficient. It can therefore be concluded, based on the observations above, that the only responsible factor for this discrepancy is the number of cells used as expected in this preliminary stage.

Conclusion

The study led to the conclusion that transient DES techniques are indeed capable of producing more accurate and physically representative results. The low cell count should be regarded as the sole reason for the lack of agreement with both experimental data and AutoCFD results. Therefore, the numerical schemes and simulation setup used in AeroCloud can be considered extendable to DES approaches as well.

The technology has thus been validated, with the only remaining requirement being an increase in the number of cells, and future implementation within AeroCloud will be possible should it prove to be of interest.

We are looking ahead to further application and testing of the technique, with the aim of evaluating its industrial implementation at AeroCloud also in view of the new seminar organized by AutoCFD: the Workshop 5, scheduled for September/October 2026.