Car design is an iterative and proprietary process. Automakers can spend several years in the design phase of a car, tweaking 3D forms in simulations, before developing the most promising designs for physical testing. The details and specifications of these tests, including the aerodynamics of a given car design, are usually not made public. Significant progress in performance, for example in fuel efficiency or electric vehicle range, may therefore be ponderous and isolated from one company to another.
MIT engineers say the search for better car designs could accelerate exponentially with the apply of generative artificial intelligence tools that can sift through massive amounts of data in seconds and find connections to generate a novel design. While such AI tools exist, the data they would need to learn from is not available, at least in any accessible, centralized form.
But now engineers have made such a dataset publicly available for the first time. The dataset, called DrivAerNet++, includes over 8,000 car designs that engineers have generated based on the most popular car types in the world today. Each design is presented in 3D and includes information about the car’s aerodynamics – how air will flow around the design, based on the group’s fluid dynamics simulations for each design.
Source: Courtesy of Mohamed Elrefaie
Each of the dataset’s 8,000 designs is available in several representations, such as a mesh, point cloud, or a straightforward list of design parameters and dimensions. Therefore, the data set can be used by various artificial intelligence models adapted to process data in a specific mode.
DrivAerNet++ is the largest open-source automotive aerodynamics dataset ever developed. Engineers envision it being used as a enormous library of realistic car designs with detailed aerodynamic data that can be used to quickly train any AI model. These models can then equally quickly generate pioneering designs that could potentially lead to more fuel-efficient cars and longer-range electric vehicles, in a fraction of the time currently needed in the automotive industry.
“This dataset lays the foundation for the next generation of artificial intelligence applications in engineering, promoting efficient design processes, lowering research and development costs, and driving progress towards a more sustainable automotive future,” says Mohamed Elrefaie, a mechanical engineering graduate student at MIT.
Elrefaie and his colleagues will present a paper detailing the novel dataset and the artificial intelligence methods that can be applied to it at the NeurIPS conference in December. Its co-authors are Faez Ahmed, assistant professor of mechanical engineering at MIT, along with Angela Dai, professor of computer science at the Technical University of Munich, and Florin Marar of BETA CAE Systems.
Filling the data gap
Ahmed directs the Design Computing and Digital Engineering (DeCoDE) Laboratory at MIT, where his group explores ways to apply artificial intelligence and machine learning tools to improve the design of sophisticated engineering systems and products, including automotive technology.
“Often, the car design process is so expensive that manufacturers can only slightly modify the car from one version to the next,” says Ahmed. “But if you have larger datasets where you know the performance of each design, you can now train machine learning models to iterate quickly, so you’re more likely to end up with a better design.”
And speed, especially in the context of advances in automotive technology, is especially pressing right now.
“This is the best time to accelerate car innovation because cars are one of the world’s biggest polluters, and the faster we can reduce this impact, the more we can help the climate,” says Elrefaie.
By looking at the novel car design process, researchers found that while there are artificial intelligence models that can rework many car designs to generate optimal designs, the actual data available for cars is constrained. Some researchers have previously collected miniature datasets of simulated car designs, while automakers rarely publish specifications for the actual designs they research, test and ultimately produce.
The team sought to fill the data gap, particularly regarding a car’s aerodynamics, which plays a key role in determining an electric vehicle’s range, and the fuel efficiency of an internal combustion engine. They realized that the challenge was to assemble a dataset of thousands of car designs, each of which was physically exact in function and form, without having to physically test and measure their performance.
To build a dataset of car designs containing a physically exact representation of their aerodynamics, researchers started with several baseline 3D models provided by Audi and BMW in 2014. These models represent three main categories of passenger cars: fastback (swallow-end sedans), notchback (sedan or a coupe with a slightly lowered rear profile) and a station wagon (such as a station wagon with a blunter, flatter rear end). Entry-level models are considered to bridge the gap between straightforward designs and more sophisticated proprietary designs, and have been used by other groups as a starting point for exploring novel car designs.
Car library
In the novel study, the team applied a transformation operation to each of the basic car models. This operation systematically resulted in miniature changes to each of 26 parameters of a given car design, such as its length, chassis features, windshield rake, and wheelbase, which were then marked as a separate car design and added to the growing dataset. Meanwhile, the team ran an optimization algorithm to ensure that each novel design would actually be distinct and not a copy of an already generated design. They then translated each 3D design in various ways so that a given design could be represented as a mesh, a point cloud, or a list of dimensions and specifications.
The researchers also performed sophisticated computational fluid dynamics simulations to calculate how air would flow around each generated car design. Ultimately, this effort resulted in over 8,000 different, physically exact 3D car models, covering the most common types of passenger cars on the road today.
To create this comprehensive dataset, researchers spent over 3 million CPU hours using MIT SuperCloud and generated 39 terabytes of data. (For comparison, it is estimated that the entire printed collection of the Library of Congress would contain approximately 10 terabytes of data.)
Engineers say researchers can now apply the dataset to train a specific artificial intelligence model. For example, an AI model can be trained on part of the dataset to learn car configurations with specific desired aerodynamics. Within seconds, the model can then generate a novel car design with optimized aerodynamics, based on lessons learned from thousands of physically exact designs in the dataset.
The researchers say the dataset can also be used for the opposite purpose. For example, after training an AI model on a dataset, designers can feed the model a specific car design and quickly estimate the design’s aerodynamics, which can then be used to calculate the car’s potential fuel economy or electric range – all without having to do the costly building and testing of a physical car.
“This dataset makes it possible to train generative AI models to perform tasks in seconds, not hours,” says Ahmed. “These models can help reduce the fuel consumption of combustion vehicles and extend the range of electric cars, ultimately paving the way for more sustainable and environmentally friendly vehicles.”
This work was supported in part by the German Academic Exchange Service and the MIT Department of Mechanical Engineering.