Artificial intelligence algorithms may be able to beat grandmasters at chess, defeat champion players at Jeopardy!, and vanquish the world’s best Go players. One thing they can’t do: Beat the world’s zippiest cyclist for the speed record of thefastest human-powered vehicle at 83.13 mph.After all, the record includes the word “human” in its name.
That may be true, but it doesn’t mean that A.I. can’t play a major role in helping design the world’s most streamlined bike by honing the necessary measurements to cut air resistance to an absolute minimum. This is what a new machine learning algorithm created at the Computer Vision Laboratory of Switzerland’s cole Polytechnique Fdrale de Lausanne (EPFL) is currently being used for. According to its creators, the algorithm was trained on the aerodynamic qualities of different 3D shapes and, as a result, has an understanding about the laws of physics that would put many flesh-and-blood designers to shame.
“Instead of computing solutions of equations or simulating moving particles, our algorithms predict the aerodynamic performances from previous experience, the same way a human engineer would do,” researcher Pierre Baqu told Digital Trends. “By doing so, we reduce the time to estimate the performance of a new design from several hours to a few milliseconds, which lets us implement computer-based automatic shape optimization.”
The technology has been spun-out as its own company, called Neural Concept, of which Baqu is CEO. It is now being used as the starting point for a newly designed “aero speed bicycle,” being built at France’s Annecy University Institute of Technology. This bike is due to have its first trials in the near-future, and will then be used (by a human) to attempt the World Human-Powered Speed Challenge in Nevada this September.
While current speed record holder Sebastiaan Bowier’s rocket-like bike looked about as optimized as is possible, it will be fascinating to see what Neural Concept’s algorithm is cable to come up with. Especially since, as Baqu has noted, it is sometimes capable of coming up with solutions that are 5 to 20 percent more aerodynamic than conventional methods.
As for the future of Neural Concept? “Our startup is developing commercial applications of the technologies for generative design, based on deep learning,” Baqu said. “We are starting collaborations with industrials, while continuously developing our software.”