This paper discusses an emerging area of applying machine 1 learning (ML) methods to augment traditional Computational 2 Fluid Dynamics (CFD) simulations of road vehicle aerodynam-3 ics. ML methods have the potential to both reduce the com-4 putational effort to predict a new geometry or car condition 5 and to explore a greater number of design parameters with the 6 same computational budget. Similar to traditional CFD meth-7 ods, there exists a broad range of approaches. In particular, 8 the accuracy and computational efficiency of a CFD simula-9 tion vary greatly depending on the choice of turbulence model 10 (DNS, LES, RANS) and the underlying spatial and temporal nu-11 merical discretizations. Similarly, the end-user must select the 12 correct ML method depending on the use-case, the available in-13 put data, and the trade-off between accuracy and computational 14 cost. In this paper, we showcase several case studies using var-15 ious data-driven ML methods to highlight the promise of these 16 approaches. Whilst these case studies are not comprehensive 17 investigations of the underlying methods and do not include all 18 possible ML approaches (i.e., physics-driven), they highlight 19 the ability of these models to in general predict new designs in 20 near real-time (i.e., less than 5 seconds), after typically less than 21 1 hour of training on a single GPU. There still exists a need for 22 high quality training data from traditional CFD methods and 23 high-fidelity CFD simulations to validate the ML predictions. 24
2023
Performance Study of Convolutional Neural Network Architectures for 3D Incompressible Flow Simulations
Ekhi Ajuria Illarramendi , Michael Bauerheim , Neil Ashton , and 2 more authors
In Proceedings of the Platform for Advanced Scientific Computing Conference, PASC 2023 , Jun 2023
Recently, correctly handling spatial information from multiple scales has proven to be essential in Machine Learning (ML) applications on Computational Fluid Dynamics (CFD) problems. For these type of applications, Convolutional Neural Networks (CNN) that use Multiple Downsampled Branches (MDBs) to efficiently encode spatial information from different spatial scales have proven to be some of the most successful architectures. However, not many guidelines exist to build these architectures, particularly when applied to more challenging 3D configurations. Thus, this work focuses on studying the impact of the choice of the number of down-sampled branches, accuracy and performance-wise in 3D incompressible fluid test cases, where a CNN is used to solve the Poisson equation. The influence of this parameter is assessed by performing multiple trainings of Unet architectures with varying MDBs on a cloud-computing environment. These trained networks are then tested on two 3D CFD problems: a plume and a Von Karman vortex street at various operating points, where the solution of the neural network is coupled to a nonlinear advection equation.
Summary of the 4th High-Lift Prediction Workshop Hybrid RANS/LES Technology Focus Group
Neil Ashton , Paul Batten , Andrew Cary , and 1 more author
This paper summarizes the collective efforts of multiple teams that contributed to the hybrid RANS/LES technical focus group for the 4th AIAA CFD High Lift Prediction Workshop (HLPW-4), which took place on January 7, 2022, in San Diego, California. The overall conclusion is that turbulence-resolving methods such as hybrid RANS/LES (HRLES) do offer improved predictions for these high-lift geometries, with respect to the underlying RANS models, but there are nuances, and some unresolved issues remain that should be the focus of future work. In particular, while HRLES methods appear to show clearly improved predictions at higher angles of attack, there is some tendency for HRLES methods to return slightly worse moment predictions at lower angles of attack, suggesting that prediction of the shallow separation from the flaps might need further research. Computing cost also remains a significant issue, with HRLES methods requiring roughly nine times more high-performance computing central processing unit core hours than steady-state RANS methods, indicating that future algorithmic and computational optimization could be beneficial. Finally, there are strong indications that modeling the wind tunnel has a positive impact on correlation with experimental measurements, suggesting that future work might be better focused on in-tunnel simulations.
2022
Towards best-practices for Hybrid RANS-LES simulations of high-lift aircraft geometries
Neil Ashton , Scott Eberhardt , Paul Batten , and 1 more author
The 2nd Automotive CFD Prediction workshop (AutoCFD2) was organized to improve the state-of-the-art in automotive aerodynamic prediction. It is the mission of the workshop organizing committee to drive the development and validation of enhanced CFD methods by establishing publicly available standard test cases for which high quality on- and off-body wind tunnel test data is available. This paper reports on the AutoCFD2 workshop for the Ford DrivAer test case. Since its introduction, the DrivAer quickly became the quasi-standard for CFD method development and correlation. The Ford DrivAer has been chosen due to the proven, high-quality experimental data available, which includes integral aerodynamic forces, 209 surface pressures, 11 velocity profiles and 4 flow field planes. For the workshop, the notchback version of the DrivAer in a closed cooling, static floor test condition has been selected. For a better comparability of CFD results, two carefully designed control meshes were provided. Both meshes share identical distributions in the flow field volume but differ in near wall spacing to allow for wall-modelled and wall-resolved solutions. The 65 results, which were submitted by 22 participants, revealed a very significant variability of the aerodynamic force predictions even when using the same turbulence model on the control grids. While individual simulations using scale-resolving hybrid turbulence models correlated very well to the experimental flow field data, other analyses using almost identical simulation approaches resulted in very different predictions. The comparison of transient versus steady state analysis confirmed that transient simulations deliver more accurate flow field predictions. A significant impact of the near wall mesh resolution could not be confirmed by the results submitted for the DrivAer test case.
Overview and Summary of the First Automotive CFD Prediction Workshop: DrivAer Model
Neil Ashton , and William Van Noordt
SAE International Journal of Commercial Vehicles, Aug 2022
The First Automotive CFD Prediction Workshop was held in December 2019 at St Anne’s College at the University of Oxford with the aim to assess the ability of a broad range of computational fluid dynamics (CFD) methods to predict the flow over realistic automotive geometries. Here, results from 53 simulation data sets from 9 separate groups are analyzed for the open-source automotive DrivAer model (in the fastback and estate variants). The represented CFD approaches include Reynolds-averaged Navier-Stokes (RANS) approaches with a broad range of turbulence models, as well as scale-resolving approaches such as wall-modelled large-eddy simulation (WMLES) and hybrid RANS-LES methods (HRLM). A range of CFD codes was used, including commercial, academic, and open source. Compared to the two experimental data points, there was a large spread of CFD results. The difference between drag predictions among HRLM and RANS methods is significant, with an even larger mismatch for lift. The differences are found to be more significant for the estate geometry than for the fastback, with the former having larger areas of flow separation. In general it is found that the spread of HRLM is smaller than those for RANS approaches, with HRLM grouping closer to the range of experimental values. However, for HRLM, there is a systematic underprediction of the front lift coefficient that is irrespective of the mesh, turbulence model, and CFD code. Given that the majority of participants used the same mesh and boundary conditions, and in some cases the same CFD code, it suggests that also user choices around numerical schemes, convergence, and turbulence model coefficients may have a sizable impact, which was not possible to fully control in this first workshop. It is worth noting as well that the CFD simulations were conducted in a free-air environment and did not model the wind-tunnel geometry itself, which may also be an area requiring further study. The DrivAer model exhibits numerous complex flow physics, i.e., laminar/turbulent separation, diffusion of momentum in turbulent shear layers, and interactions of turbulent wakes with boundary layers. However, the nature of a community-driven workshop and the lack of extensive experimental data means that this article can only report the current state of the art and serve as a reference for future workshops and a springboard for more focused future studies where topics can be explored in greater detail.
HLPW-4/GMGW-3: Hybrid RANS/LES Technology Focus Group Workshop Summary
N Ashton , P Batten , A Cary , and 1 more author
In AIAA Aviation 2022 , Aug 2022
2021
Performance of cpu and gpu hpc architectures for off-design aircraft simulation
Rotating and swirling turbulence comprises an important class of turbulent flows, not only due to the complex physics that occur, but also due to their relevance to many engineering applications, such as combustion, cyclone separation, mixing, etc. In these types of flows, rotation strongly affects the characteristics and structure of turbulence. The underlying turbulent flow phenomena are complex and currently not well understood. The axially rotating pipe flow is a well-suited prototypical case for studying rotation effects in turbulence due to its simple geometry and the ability to be reproduced experimentally in a controlled environment. By examining the complex interaction of turbulent structures within rotating turbulent pipe flow, insight can be gained into the behavior of rotating flows relevant to engineering applications. Direct numerical simulations are conducted at a bulk Reynolds number of ReD = 19,000 with rotation numbers ranging from N = 0 to 3. In addition to providing turbulence statistics, proper orthogonal decomposition is used to identify the relevant (highest energy) modes of the flow and obtain an understanding about the coherence in the flow.
Towards High-Fidelity CFD on the Cloud for the Automotive and Motorsport sectors
New direct numerical simulation data of a fully-developed axially rotating pipe at Re = 5300 and Re = 19, 000 is used to examine the performance of the second-moment closure elliptic blending Reynolds stress model for a range of rotation rates from N=0 to N=3. In agreement with previous studies (using alternative second-moment closure models), the turbulence suppression observed by the DNS is over-predicted. This over-prediction is greatest at Re = 5, 300 and most noticeable in the poor prediction of the u′ w′ turbulent shear-stress component. At N=3 the flow is completely relaminarized in contrast to the DNS that is only partly relaminarized. The accuracy of the second-moment closure model is superior to the two-equation k − ω SST model which predicts pure solid-body rotation, however, both are equally poor at the highest rotation rates. The accuracy of each model is also assessed for the initial portion of a rotating pipe where in contrast to the fully-developed rotating pipe flow the turbulent suppression is under-predicted compared to the DNS. It is clear that greater work is required to understand the root cause of the poor prediction by these second-moment closure models and further DNS and experimental work is underway to assist this effort.
A DNS Study to Investigate Turbulence Suppressionin Rotating Pipe Flows
J Davis , G Sparsh , N Ashton , and 2 more authors
In , Jun 2019
2018
Real-World Variability in the Prediction of Intracranial Aneurysm Wall Shear Stress: The 2015 International Aneurysm CFD Challenge
Kristian Valen-Sendstad , Aslak W. Bergersen , Yuji Shimogonya , and 52 more authors
Cardiovascular Engineering and Technology, Jun 2018
Assessing the Sensitivity of Hybrid RANS-LES Simulations to Mesh Resolution, Numerical Schemes and Turbulence Modelling within an Industrial CFD Process
With the increased availability of computational resources, the past decade has seen a rise in the use of computational fluid dynamics (CFD) for medical applications. There has been an increase in the application of CFD to attempt to predict the rupture of intracranial aneurysms, however, while many hemodynamic parameters can be obtained from these computations, to date, no consistent methodology for the prediction of the rupture has been identified. One particular challenge to CFD is that many factors contribute to its accuracy; the mesh resolution and spatial/temporal discretization can alone contribute to a variation in accuracy. This failure to identify the importance of these factors and identify a methodology for the prediction of ruptures has limited the acceptance of CFD among physicians for rupture prediction. The International CFD Rupture Challenge 2013 seeks to comment on the sensitivity of these various CFD assumptions to predict the rupture by undertaking a comparison of the rupture and blood-flow predictions from a wide range of independent participants utilizing a range of CFD approaches. Twenty-six groups from 15 countries took part in the challenge. Participants were provided with surface models of two intracranial aneurysms and asked to carry out the corresponding hemodynamics simulations, free to choose their own mesh, solver, and temporal discretization. They were requested to submit velocity and pressure predictions along the centerline and on specified planes. The first phase of the challenge, described in a separate paper, was aimed at predicting which of the two aneurysms had previously ruptured and where the rupture site was located. The second phase, described in this paper, aims to assess the variability of the solutions and the sensitivity to the modeling assumptions. Participants were free to choose boundary conditions in the first phase, whereas they were prescribed in the second phase but all other CFD modeling parameters were not prescribed. In order to compare the computational results of one representative group with experimental results, steady-flow measurements using particle image velocimetry (PIV) were carried out in a silicone model of one of the provided aneurysms. Approximately 80% of the participating groups generated similar results. Both velocity and pressure computations were in good agreement with each other for cycle-averaged and peak-systolic predictions. Most apparent "outliers" (results that stand out of the collective) were observed to have underestimated velocity levels compared to the majority of solutions, but nevertheless identified comparable flow structures. In only two cases, the results deviate by over 35% from the mean solution of all the participants. Results of steady CFD simulations of the representative group and PIV experiments were in good agreement. The study demonstrated that while a range of numerical schemes, mesh resolution, and solvers was used, similar flow predictions were observed in the majority of cases. To further validate the computational results, it is suggested that time-dependent measurements should be conducted in the future. However, it is recognized that this study does not include the biological aspects of the aneurysm, which needs to be considered to be able to more precisely identify the specific rupture risk of an intracranial aneurysm.
Slat Noise Prediction using Hybrid RANS-LES methods on Structured and Unstructured Grids
Neil Ashton , Alastair West , and Fred Mendonca
In 21st AIAA/CEAS Aeroacoustics Conference , Jun 2015
Key factors in the use of DDES for the flow around a simplified car
N Ashton , and A Revell
International Journal of Heat and Fluid Flow, Jun 2015
2013
Development of an Alternative Delayed Detached-Eddy Simulation Formulation Based on Elliptic Relaxation
N Ashton , A Revell , R Prosser , and 1 more author
AIAA Journal, Jun 2013
2012
Embedded DDES of 2D Hump Flow
R Poletto , A Revell , T Craft , and 1 more author
Jun 2012
2011
A hybrid numerical scheme for a new formulation of delayed detached-eddy simulation (DDES) based on elliptic relaxation
N Ashton , R Prosser , and A Revell
In Journal of Physics: Conference Series , Dec 2011