Do you want to design the car of the future? Here are 8,000 designs to get you started. | MIT News

Car design is an iterative and proprietary process. Automakers can spend several years in the design phase of a car, optimizing 3D shapes in simulations before developing the most promising designs for physical testing. The details and specifications of these tests, including the aerodynamics of a particular vehicle design, are not typically made public. Significant performance advances, such as in fuel efficiency or the range of electric vehicles, can therefore occur slowly and in isolation from company to company.

MIT engineers say the search for better car designs can be accelerated exponentially by using generative artificial intelligence tools that can sift through massive amounts of data in seconds and find connections to generate a novel design. Although such AI tools exist, the data they would need to learn from has not been available, at least not in any accessible, centralized form.

But now the engineers have made such a data set available to the public for the first time. The data set, called DrivAerNet++, includes more than 8,000 vehicle designs that engineers created based on today’s most common vehicle types in the world. Each design is presented in 3D form and includes information about the car’s aerodynamics – the way air would flow around a particular design, based on fluid dynamics simulations the group conducted for each design.

Each of the 8,000 drafts of the dataset is available in different representations, such as: B. as a mesh, point cloud or as a simple list of the parameters and dimensions of the design. Therefore, the data set can be used by different AI models that are tuned to process data in a specific modality.

DrivAerNet++ is the largest open source vehicle aerodynamics dataset developed to date. The engineers envision an extensive library of realistic vehicle designs with detailed aerodynamic data that can be used to quickly train any AI model. These models can then just as quickly produce novel designs that could potentially lead to more fuel-efficient cars and longer-range electric vehicles, in a fraction of the time it takes the automotive industry today.

“This data set lays the foundation for the next generation of AI applications in engineering, promoting efficient design processes, reducing R&D costs, and driving progress toward a more sustainable automotive future,” says Mohamed Elrefaie, a graduate student in mechanical engineering at MIT.

Elrefaie and his colleagues will present a paper detailing the new data set and AI methods that could be applied to it at the NeurIPS conference in December. His co-authors are Faez Ahmed, assistant professor of mechanical engineering at MIT, along with Angela Dai, associate professor of computer science at the Technical University of Munich, and Florin Marar of BETA CAE Systems.

Closing the data gap

Ahmed leads the Design Computation and Digital Engineering Lab (DeCoDE) at MIT, where his group explores ways to use AI and machine learning tools to improve the design of complex engineering systems and products, including automotive technology.

“When designing a car, the forward process is often so expensive that manufacturers can only tweak a car a little from one version to the next,” says Ahmed. “But if you have larger data sets where you know the performance of each design, you can now train machine learning models to iterate quickly, so you are more likely to get a better design.”

And speed, particularly in advancing automotive technology, is particularly urgent now.

“This is the best time to drive car innovation because cars are one of the biggest polluters in the world and the faster we can reduce this contribution, the more we can help the climate,” says Elrefaie.

When looking at the process of new car design, the researchers found that while there are AI models that could run through many car designs to generate optimal designs, the actual car data available is limited. Some researchers had previously compiled small data sets of simulated car designs, while automakers rarely release the specifications of the actual designs they study, test and ultimately produce.

The team wanted to fill the data gap, particularly when it comes to a car’s aerodynamics, which plays a key role in determining the range of an electric vehicle, and the fuel efficiency of an internal combustion engine. They realized that the challenge was to assemble a data set of thousands of car designs, each physically correct in function and form, without the benefit of having to physically test and measure their performance.

To create a dataset of car designs with physically accurate representations of their aerodynamics, researchers started with several basic 3D models provided by Audi and BMW in 2014. These models represent three main categories of passenger cars: Fastback (sedans with a sloping rear). end), notchback (sedans or coupes with a slight slope in the rear profile) and station wagon (e.g. station wagons with blunter, flatter hedges). The base models are intended to bridge the gap between simple designs and more complicated proprietary designs and have been used by other groups as a starting point for researching new car designs.

Car library

In their new study, the team applied a morphing operation to each of the base car models. This process systematically made minor changes to each of the 26 parameters of a given vehicle design, such as length, underbody features, windshield slope, and wheel arch, which were then labeled as a distinct vehicle design and then added to the growth data set. Meanwhile, the team ran an optimization algorithm to ensure that each new design was actually unique and not a copy of an already generated design. They then translated each 3D design into different modalities so that a given design can be represented as a mesh, point cloud, or list of dimensions and specifications.

The researchers also conducted complex computational fluid dynamics simulations to calculate how air would flow around each generated car design. Ultimately, these efforts created more than 8,000 different, physically accurate 3D car shapes, spanning the most common types of passenger vehicles on the road today.

To create this comprehensive data set, researchers spent over 3 million CPU hours on the MIT SuperCloud and generated 39 terabytes of data. (For comparison, it is estimated that the Library of Congress’s entire print collection would contain approximately 10 terabytes of data.)

The engineers say researchers can now use the data set to train a specific AI model. For example, an AI model could be trained on a portion of the data set to learn vehicle configurations that have specific desired aerodynamics. Within seconds, the model could then generate a new car design with optimized aerodynamics based on insights from the dataset’s thousands of physically correct designs.

The researchers say the data set could also be used for the opposite goal. For example, after training an AI model on the data set, designers could feed the model a specific car design and have it quickly estimate the design’s aerodynamics, which can then be used to calculate the car’s potential fuel efficiency or electric range – and that all without carrying out expensive building and testing of a physical car.

“With this data set, you can train generative AI models to do things in seconds instead of hours,” says Ahmed. “These models can help reduce the fuel consumption of internal combustion engine vehicles and increase the range of electric cars – ultimately paving the way for more sustainable, environmentally friendly vehicles.”

This work was supported in part by the German Academic Exchange Service and the Department of Mechanical Engineering at MIT.