Neil Ashton

Driven by the advancement of GPUs and AI, the field of Computational Fluid Dynamics (CFD) is undergoing significant transformations. This paper bridges the gap between the machine learning and CFD communities by deconstructing industrial-scale CFD simulations into their core components. Our main contribution is to propose the first scaling law that incorporates CFD inputs for both data generation and model training to outline the unique challenges of developing and deploying these next-generation AI models for complex fluid dynamics problems. Using our new scaling law, we establish quantitative estimates for the large-scale limit, distinguishing between regimes where the cost of data generation is the dominant factor in total compute versus where the cost of model training prevails. We conclude that the incorporation of high-fidelity transient data provides the optimum route to a foundation model. We constrain our theory with concrete numbers, providing the first public estimates on the computational cost and time to build a foundation model for CFD.

In this paper, we introduce a benchmarking framework within the open-source NVIDIA PhysicsNeMo-CFD framework designed to systematically assess the accuracy, performance, scalability, and generalization capabilities of AI models for automotive aerodynamics predictions. The open extensible framework enables incorporation of a diverse set of metrics relevant to the Computer-Aided Engineering (CAE) community. By providing a standardized methodology for comparing AI models, the framework enhances transparency and consistency in performance assessment, with the overarching goal of improving the understanding and development of these models to accelerate research and innovation in the field. To demonstrate its utility, the framework includes evaluation of both surface and volumetric flow field predictions on three AI models: DoMINO, X-MeshGraphNet, and FIGConvNet using the DrivAerML dataset. It also includes guidelines for integrating additional models and datasets, making it extensible for physically consistent metrics. This benchmarking study aims to enable researchers and industry professionals in selecting, refining, and advancing AI-driven aerodynamic modeling approaches, ultimately fostering the development of more efficient, accurate, and interpretable solutions in automotive aerodynamics

Machine Learning (ML) has the potential to revolutionise the field of automotive aerodynamics, enabling split-second flow predictions early in the design process. However, the lack of open-source training data for realistic road cars, using high-fidelity CFD methods, represents a barrier to their development. To address this, a high-fidelity open-source (CC-BY-SA) public dataset for automotive aerodynamics has been generated, based on 500 parametrically morphed variants of the widely-used DrivAer notchback generic vehicle. Mesh generation and scale-resolving CFD was executed using consistent and validated automatic workflows representative of the industrial state-of-the-art. Geometries and rich aerodynamic data are published in open-source formats. To our knowledge, this is the first large, public-domain dataset for complex automotive configurations generated using high-fidelity CFD.

The development of Machine Learning (ML) methods for Computational Fluid Dynamics (CFD) is currently limited by the lack of openly available training data. This paper presents a new open-source dataset comprising of high fidelity, scale-resolving CFD simulations of 500 geometric variations of the Ahmed Car Body - a simplified car-like shape that exhibits many of the flow topologies that are present on bluff bodies such as road vehicles. The dataset contains simulation results that exhibit a broad set of fundamental flow physics such as geometry and pressure-induced flow separation as well as 3D vortical structures. Each variation of the Ahmed car body were run using a high-fidelity, time-accurate, hybrid Reynolds-Averaged Navier-Stokes (RANS) - Large-Eddy Simulation (LES) turbulence modelling approach using the open-source CFD code OpenFOAM. The dataset contains boundary, volume, geometry, and time-averaged forces/moments in widely used open-source formats. In addition, the OpenFOAM case setup is provided so that others can reproduce or extend the dataset. This represents to the authors knowledge, the first open-source large-scale dataset using high-fidelity CFD methods for the widely used Ahmed car body that is available to freely download with a permissive license (CC-BY-SA).

This paper presents a new open-source high-fidelity dataset for Machine Learning (ML) containing 355 geometric variants of the Windsor body, to help the development and testing of ML surrogate models for external automotive aerodynamics. Each Computational Fluid Dynamics (CFD) simulation was run with a GPU-native high-fidelity Wall-Modeled Large-Eddy Simulations (WMLES) using a Cartesian immersed-boundary method using more than 280M cells to ensure the greatest possible accuracy. The dataset contains geometry variants that exhibits a wide range of flow characteristics that are representative of those observed on road-cars. The dataset itself contains the 3D time-averaged volume & boundary data as well as the geometry and force & moment coefficients. This paper discusses the validation of the underlying CFD methods as well as contents and structure of the dataset. To the authors knowledge, this represents the first, large-scale high-fidelity CFD dataset for the Windsor body with a permissive open-source license (CC-BY-SA).

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.

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.

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.

© 2021, American Institute of Aeronautics and Astronautics Inc, AIAA. All rights reserved. This paper presents a detailed analysis of the relative performance and cost of GPU and CPU architectures for a full aircraft RANS simulation using the CFD code zCFD. Using Amazon Web Services as the platform, several generations of NVIDIA GPUs are assessed (T4, V100, and A100) and compared to x86 Intel Broadwell and Skylake CPUs. It is found that for the same computational setup and mesh, one Amazon EC2 p4d.24xlarge instance (containing eight NVIDIA A100 GPUs) provides the same run-time as 2,160 cores (60 nodes) of the Intel Skylake Amazon EC2 c5n.18xlarge. At the maximum number of nodes tested (100 CPU nodes and 10 GPU nodes-each containing 8 GPUs) the GPU-based instance is 2.8x faster than the x86 Intel CPU-based instance (Amazon EC2 c5n.18xlarge) for half the cost. It is hoped these results will provide extra evidence to support the continued focus on CFD code development to support GPUs and will help to make high-fidelity simulations more practical for engineering companies.

This paper presents a comprehensive investigation of RANS and DES models for the Ahmed car body and a realistic automotive vehicle; the DrivAer model. A variety of RANS models, from the 1-equation Spalart Allmaras model to a low-Reynolds number Reynolds Stress model have shown an inability to consistently correctly capture the flow field for both the Ahmed car body and DrivAer model, with the under-prediction of the turbulence in the initial separated shear layer found as a key deficiency. It has been shown that the use of a hybrid RANS-LES model (in this case, Detached Eddy Simulation) offers an advantage over RANS models in terms of the force coefficients, and general flow field for both the Ahmed car body and the DrivAer model. However, for both cases even at the finest mesh level hybrid RANS-LES methods still exhibited inaccuracies. Suggestions are made on possible improvements, in particular on the use of embedded LES with synthetic turbulence generation. Finally the computational cost of each approach is compared, which shows that whilst hybrid RANS-LES offer a clear benefit over RANS models for automotive relevant flows they do so at a much increased cost.