Designing cars is a long, costly process.
Automakers spend years testing and tweaking car designs, especially to improve aerodynamics, which affects fuel efficiency and the range of electric vehicles (EVs).
However, these tests are often private, and breakthroughs happen slowly across the industry.
MIT engineers have found a way to speed things up dramatically using artificial intelligence (AI).
They’ve created a massive open-source dataset called DrivAerNet++, which contains over 8,000 car designs complete with detailed aerodynamic data.
This dataset could help engineers and designers develop fuel-efficient cars and EVs with longer ranges faster than ever before.
DrivAerNet++ includes 3D representations of car designs that fall into three main categories: fastback (sloped rear, like many sedans), notchback (a more angled rear, like a coupe), and estateback (flat rear, like station wagons).
These designs were generated by systematically tweaking features like length, windshield angle, and wheel size.
For each design, the MIT team simulated how air would flow around the car, creating highly accurate aerodynamic data.
This information is essential because aerodynamics significantly impacts fuel efficiency and the energy consumption of EVs.
To ensure variety, the team used advanced algorithms to check that no two designs were identical.
Each car model is available in different formats—such as 3D meshes, point clouds, or a list of dimensions—so researchers can use the dataset with various AI tools.
Creating this dataset required an enormous effort. The team used the MIT SuperCloud to run more than 3 million hours of computer simulations, producing 39 terabytes of data.
That’s nearly four times the size of the entire printed collection of the Library of Congress!
Traditionally, carmakers rely on physical testing, which is slow and expensive. AI models can process the DrivAerNet++ dataset and generate new car designs with optimized aerodynamics in seconds instead of hours or days. This rapid iteration could lead to faster innovation in car design.
For example, an AI model could learn from the dataset to create a design with low air resistance, which could improve fuel efficiency for gas-powered cars or extend the range of EVs. Alternatively, the dataset could be used to quickly estimate the aerodynamic performance of an existing design without building a physical prototype.
“This is the best time to accelerate car innovations,” says Mohamed Elrefaie, an MIT mechanical engineering graduate student and co-author of the study. “Cars are one of the biggest polluters, so the faster we reduce their emissions, the more we help the planet.”
Until now, AI tools for car design have been limited by the lack of available data. Car manufacturers rarely share the details of their designs and tests. While some small datasets of simulated car designs exist, they’re not large or diverse enough to train powerful AI models.
DrivAerNet++ changes that by offering a large, open-source dataset that covers a wide range of realistic car designs.
“This dataset lets AI models do in seconds what used to take hours,” says Faez Ahmed, assistant professor at MIT and co-author of the study. “It can help lower fuel consumption and increase the range of EVs, paving the way for more sustainable cars.”
The MIT team hopes DrivAerNet++ will inspire researchers and carmakers to develop greener vehicles. By making this data open-source, they aim to remove barriers to innovation and promote collaboration across the industry.
The team will present their work at the NeurIPS 2024 conference, where they’ll also showcase how AI models can use the dataset to advance car design. With tools like DrivAerNet++, the future of cars could arrive much faster—and with less impact on the environment.
