MIT engineers have set out to prove that artificial intelligence can generate significant advancements in car design, especially in the realm of aerodynamics. As in other fields, AI can comb through a massive amount of data, separate what’s useful from what’s not, and design a car accordingly. That’s not exactly great news if you’re currently studying, say, automotive design, but MIT claims AI can make cars more efficient.
The AI tools required to design a car from scratch with an eye on aerodynamics already exist, but the data was either private, decentralized, or missing altogether until recently. That’s where MIT engineers stepped in. They created a dataset called DrivAerNet++ that includes more than 8,000 different car designs represented in 3D form. Each design includes simulation-based aerodynamic information about the car it represents.
Achieving this was easier said than done. Engineers started with a handful of baseline 3D models provided by Audi and BMW in 2014 which they split into three categories labeled fastback, notchback, and estateback, respectively. (MIT is apparently leaving convertible design to humans.) With these files loaded and ready to go, the MIT team made slight changes to 26 different parameters (including the length, various underbody parts, and the slope of the windshield) and saved each modification as a separate car. That’s how they ended up with more than 8,000 designs.
We haven’t looked through the 8,000 designs, but we spot some familiar-looking cars in the few that were published by MIT. One is clearly based on an E91-generation BMW 3 Series Touring, albeit with acid trip-like colors and a handful of visual tweaks. Another looks like an Audi A4 Avant, so at least our future AI overlords will allow us to continue driving wagons (probably in exchange for a pencil and a sheet of paper as a sacrifice).
MIT hopes that engineers around the globe will use this gigantic amount of data to teach AI models how to design a car. In turn, anyone with access to these tools—including you, potentially—will be able to design a wagon faster than, say, General Motors can and theoretically at a lower cost.
“Often when designing a car, the forward process is so expensive that manufacturers can only tweak a car a little bit from one version to the next,” said Faez Ahmed, assistant professor of mechanical engineering at MIT. “But if you have larger datasets where you know the performance of each design, now you can train machine-learning models to iterate fast so you are more likely to get a better design.”
Can the industry train AI to sculpt a car that’s both aerodynamic and visually appealing? Will AI eventually be able to take into account the other factors that impact a car’s design process, such as production feasibility and regulation-mandated crumple zones? And, why the hell not: Will “pre-AI model” one day get tossed around as a selling point in the enthusiast community? Time will tell—and we may not have to wait long to find out.
Got tips? Send ’em to tips@thedrive.com