AI has been transforming many creative areas, and the car industry is one sector that's particularly excited about its potential. Creating a new car is hugely expensive and time-consuming because of the number of iterations required to first design and then prototype a vehicle.
The design of electric vehicles has often been controversial too, from the Tesla Cybertruck design fails to the bold Jaguar Type 00 concept. Only recently, we've seen the problems faced by the Cybertruck in snow. But researchers at MIT think AI could help design the perfect electric car, and they've created an open-source database.
The DrivAerNet++ database was compiled from 39 terabytes of data, consuming 3 million central processing unit hours in the MIT SuperCloud. It includes over 8,000 3D models based on the designs of the existing cars created by an algorithm that adjusted 26 parameters, from vehicle length to underbody features, windshield slope and tread and wheel shapes for each baseline model.
The team also ran an algorithm to determine whether any newly generated design was a copy of a car that already existed or a new design. The 3D designs were then converted into readable formats: a mesh, point cloud and a list of dimensions and specs. Fluid dynamics simulations were run to calculate how air would flow around each generated design in order to provide specifications on aerodynamics.
The idea is that the dataset could be used to train an AI model that would then be able to seek out the best combination of features, from aerodynamic design to an efficient and eco-friendly motor, reducing research and development costs and speeding up car design.
Faez Ahmed, assistant professor of mechanical engineering at MIT, said: "The forward process is so expensive that manufacturers can only tweak a car a little bit from one version to the next. 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."
The team described its paper at the NeurIPS conference in Vancouver in December.
