Credit: JLStock / Shutterstock
Ideas Made to Matter
Artificial intelligence can help design more appealing cars
From curb appeal in real estate to smooth edges on smartphones, consumers gravitate toward products that are pleasing to the eye. This is especially true in the automotive industry, where product aesthetics have been linked to roughly 60% of purchasing decisions.
“People buy cars based on aesthetics. Styling can make a difference,” said a professor of marketing at MIT Sloan. Styling is also expensive: Carmakers invest more than $1 billion to design the average car model and up to $3 billion for major redesigns.
A recent paper Hauser co-authored demonstrates that machine learning models can not only predict the appeal of new aesthetic designs but also generate designs that are aesthetically pleasing or aesthetically innovative. (And, once trained, the models can run on a standard corporate laptop.)
The paper was co-authored by Yale School of Management professor Alex Burnap and Kellogg School of Management professor Artem Timoshenko.
“The models are a tool for designers to get new ideas and try them out,” Hauser said. “They are capable of generating new images that are highly aesthetically pleasing and that can be evaluated quickly.”
Unappealing cars do not sell
The Pontiac Aztek is an infamous example of how car buyers prioritize aesthetics.
Product aesthetics have been linked to roughly 60% of purchasing decisions in the automotive industry.
General Motors released the Aztek in the summer of 2000, building the crossover SUV on the same platform as the Buick Rendezvous. With multiple features for fans of the outdoors, the Aztek generally earned high customer satisfaction scores — apart from its exterior styling.
Here, the Aztek flopped. A profile noted that the vehicle had an intentionally aggressive, “in your face” design and wasn’t for everyone. It has been routinely derided as one of the ugliest cars of all time, and GM stopped making the SUV in 2005.
The Aztek sold half as many units as the Rendezvous, which was subsequently redesigned and rereleased as the Buick Enclave — which sold at a 30% higher manufacturer’s suggested retail price. The Enclave is still manufactured today, more than 15 years after its initial launch.
The Aztek offers a clear lesson, Hauser said: “If two cars are equally reliable and effective, consumers will buy the one that’s more attractive.”
Using AI to predict — and generate — aesthetically pleasing models
Today’s carmakers make big investments to avoid releasing the next Aztek.
Traditionally, this process has relied on theme clinics. These are events where carmakers bring hundreds of targeted consumers to a single location to judge designs. Theme clinics can cost $100,000 each, and carmakers need to hold hundreds each year to make sure they put the right designs into production.
Here, predictive modeling has an obvious appeal: Carmakers that can weed out the designs most likely to earn low scores on aesthetics won’t bother advancing these options beyond the initial design stage. With fewer designs that need to be tested in theme clinics, development timelines will get shorter and costs will decrease.
Working with GM as a research partner, Hauser and his co-authors developed two models:
- A generative model that creates new car designs based on prompts from designers about viewpoints, colors, body type, and image.
- A predictive model that forecasts how consumers will rate designs with respect to aesthetic appeal or innovativeness.
Research began with the predictive model, built on a deep neural network. This model achieved the desired results, with a 43.5% improvement over the baseline — and an improvement over more conventional machine learning models.
“Our model was able to indicate the designs that were good and the designs that were bad,” Hauser said. “But as we got more and more into the process, we realized the real leverage was in creating new designs.”
The generative model produced images that consumers deemed to be aesthetically appealing and even suggested designs that were later introduced to the marketplace. The researchers also found that the model can be applied to nonautomotive products.
Augmenting the design experience
As is the case with other successful applications of artificial intelligence, the models aren’t meant to replace human designers. For starters, the generative model doesn’t just spit out designs automatically; it needs an experienced designer to define the parameters first, Hauser said.
In addition, automotive design is an inherently iterative and asynchronous process. Designers iterate through design-concept generation, testing, evaluation, and redesign. The finished product — an amalgamation of tens of thousands of decisions — gets rated by consumers and critics alike, based on descriptors such as sporty, rugged, luxurious, and so on.
Hauser and his co-authors view artificial intelligence as an augmentation of the design process akin to computer-assisted modeling in furniture design, fashion, and other industries where aesthetics plays a prominent role.
“There are a number of different ways you can cut a dress,” he said. “A machine learning model can give designers ideas about what customers will think is aesthetically pleasing, but a designer isn’t going to produce exactly what the machine puts out.”
Read next: Machine Learning, explained