AI/ML in AutoCFD: Benchmarking for Automotive CFD

This blog post delves into the exciting intersection of Artificial Intelligence (AI), Machine Learning (ML), and Computational Fluid Dynamics (CFD) within the context of automotive engineering. We'll explore the motivations behind integrating these technologies, discuss available datasets for benchmarking AI/ML methods in CFD, and highlight the importance of transparent and consistent assessment for new approaches. This is a crucial area for automotive engineers looking to leverage cutting-edge techniques for improved designs and simulations.

The Rise of AI/ML in CFD

Over the past five to seven years, research and development in Machine Learning and Artificial Intelligence (AI/ML) for CFD have significantly accelerated.

This trend has been further amplified in the last two years, driven by the emergence of ML4CAE startups and the exponential growth in AI/ML since the launch of ChatGPT in November 2022.

This surge in AI/ML adoption within CFD presents a wide array of options for the automotive CFD community. Many startups and ISVs are now offering deep-learning-based, data-driven, or physics-driven surrogate models, including well-known offerings such as Altair PhysicsAI, Ansys SimAI, BeyondMath, and Nvidia Modulus. These packages and frameworks provide automotive engineers with various paths to explore the benefits of AI/ML in their workflows.

AutoCFD series has helped new and existing CFD codes to benchmark their methods and explore new approaches. We want AutoCFD to also play a central role in the transparent and consistent way of assessing new AI/ML methods for automotive CFD.

The Role of AutoCFD in Assessing New AI/ML Methods

The AutoCFD series has historically played a vital role in helping new and existing CFD codes benchmark their methodologies and explore innovative approaches. With the growing prominence of AI/ML in CFD, AutoCFD is poised to take on an even more crucial function: to provide a central hub for the transparent and consistent evaluation of new AI/ML methods tailored for automotive CFD applications.

Assessing these methods fairly and consistently is critical for progress and for building trust in AI/ML-driven CFD solutions. AutoCFD, in this capacity, helps to ensure that new approaches are rigorously tested and validated before widespread adoption.

The end goal is to create a standardized and reliable testing ground for evaluating the efficacy of new AI/ML methods, fostering innovation and accelerating the development of more efficient and accurate simulation tools for the automotive industry.

How ChatGPT Accelerated AI in Automotive

The release of ChatGPT in November 2022 acted as a seminal moment. It catalysed global interest and a significant increase in investment into the field of AI, especially into the Large Language Models space. VC funding saw the massive potential Generative AI and its AI uses have in the large language model space and the potential for it in the CAE space.

Navigating and Utilizing the Data Sets

The speaker goes over the importance of data and how this can be used to implement new AI features.

It must be shared to achieve open results. It is because of this that test cases are used that are not the company’s own car, and this keeps things open for transparency. They’ve also realized a couple of issues for training data and generating 3D outputs.

These items are generally accepted that with current technologies, the models will not replace high fidelity CFD, and will need to create training data so that the models may be used as an additional rapid conceptual design tool. That's that all software companies who have submitted abstracts withdrew their abstracts

It is impossible to properly one on one, he appreciates that this audience is a mix of people highly familiar with AI, and others completely new.

In general he discusses surrogate models where inputs CFD data and metrics and these data sets would need to be on HAND to do these things.

Understanding Data Structures

To promote a clear comprehension and efficient utilization of these complex datasets, the team has adopted a structured approach. Specifically, each dataset incorporates a consistent naming convention for all files and components.

This helps so that people sharing openly can give the best results.

Key elements within each dataset include:

• STL files: These files define the surface mesh of the Ahmed car body geometry.
• Boundary VTP: These files describe the time-averaged flow quantities (Pressure Coefficient, Skin Friction Coefficient, y+) on the Ahmed car body surface.
• Volume VTU: Files detail the time-averaged flow quantities (pressure, velocity, Reynolds Stresses, Turbulent Kinetic Energy) within the domain volume.
• Force/Moment CSV: Files contain time-averaged drag & lift and side force and pitching moment coefficients.
• Image Slices: Folder with the various images that you can utilize, as well as the STL and .csv files.

Accessing the Data

The speaker goes over the logistics, with clear data labels and data structures, such as .stl, VTP, .CSV, VTD, images, as well as OpenFoam setup to extend and reproduce the data.

These files are on Amazon S3 without needing an Amazon web services account, free to download. In addition, he lists a website caemldatasets.org/ahmedml for all the details, or pre-print and commercial usage licenses. You can view all three data sets below.

Cost Consideration for GPU and Data

As this technology is in the beginning stages, costs can be high, especially with all of the compute resources that you require.

At a cost of a cent, however, 500 case runs adds up. He makes sure to say that in many machine learning cases the training model will be used. It also makes the point that there will need to be an understanding of GPU before you can get started.

The Main Goal of AI Integration in CFD

Overall, the main idea is that it looks at the difference between a geometry flow and its potential relationships and it should all be automated. Also he states that there needs to be easy and clear validation standards.

Application in Turbulence models

Many of the uses of Machine Learning in CFD could apply to a turbulence model, doing something with an initialization, and many linear solver or transition models.

But, more specifically, they are most interested in the surrogate models.