Machine Learning Software Monolith Cuts Development Time for BMW’s Test Engineers
Monolith AI focused on building software that allows engineering teams to work with historical test data easily, without the need for constant consultation with their data-science experts.
BMW faced a problem. Because leg injuries can be common in vehicle crashes, the effect of all manner of deformation must be measured in order to predict, with a mathematical formula called the tibia index that predicts how likely knee injuries are in a crash. That’s required by global regulating bodies including the NHTSA in the U.S. But doing so led the manufacturer to a place at which physical test results were not very predictable because of the many variables involved.
So BMW worked with London-based Monolith AI to leverage decades of existing test data to use machine learning to detect patterns that allowed them to more accurately predict and reduce the likelihood of the leg injuries indicated by the tibia index.
Monolith’s founder and CEO, Richard Ahlfeld, joined Monolith’s marketing team at the 2022 Automotive Testing Expo in Novi, Mich., to show the potential of the company’s software to let vehicle engineers “test less and learn more,” using the software tool, which was built so it doesn’t require a specialist in data science to use.
“We looked at how artificial intelligence or machine learning can impact product development tests,” Ahlfeld said. “We’ve looked at gas flow meters with Honeywell, turbines and EVs for Siemens and aircraft engines with RollsRoyce."
Ultimately, he said the company asked, “How do we not come in as a consultant, but how do we build software so people can do this? This is what we did.” Ahlfeld said the company invested $12 million into its Monolith platform with which engineers can “do crash tests, for example, by [themselves] without asking anyone for help.”
BMW’s good problem? A wealth of data
BMW’s engineers were being challenged by the tibia index problem. A Monolith white paper on the project says that standard crash test methods and even simulations had trouble achieving high correlation with real-world crash data. And the process was so time consuming that doesn’t fit with today’s faster-to-market development demands.
“It takes approximately seven years to plan, develop, test, pre-build and begin production of a new car from scratch,” said Oliver Walter, former BMW product manager for the 7 Series and the i3 and now a consultant.
What BMW did have, Ahlfeld said, was decades of testing data that they didn’t know how to apply to the problem. “Testing creates a lot of unused data,” he said, citing a hypothetical crash example: In a test having 1,000 different parameters [not unusual for a crash test] achieved by 30,000 components hitting a wall at 60 mph (96.5 km/h) in different conditions with 1,300 sensors on the car with a sounding rate of 100 Hz per second, it creates 1.4 gigabytes of data every second.
Whether it’s a “crash test database at BMW or an NVH database at Siemens, that realization — that unused R&D data are extremely valuable — is what created Monolith, Ahlfeld said. There were people who had been with BMW asking why so many physical tests had to be repeated. “So, we try to make this data more useful to test engineers so they can learn more from it.”
For BMW, the Monolith team helped the engineers use machine learning to uncover patterns. “One crash test may look like random noise,” he said, “but if you run pattern recognition over 10,000 tests you can figure out the problem and solve it, meaning BMW can achieve a [NHTSA] 5-star safety rating and, as a result, sell more cars.” He said it shaved months off the normal development time because they could do fewer full-size prototype crash tests. “It can cut development time by as much as 50 percent.”
Ahlfeld said machine learning is not the answer to every engineering problem, because simulation and real-world testing work in many instances. He said it is most useful when engineers are confronted with “intractable physics problems,” which he defined as situations in which one can’t generate a simplified physical model that describes all the complications experienced in a prototype.
Engineering? Meet data science
Ahlfeld said that in many cases, unchanging engineering methods and hiring practices mean that engineers aren’t data scientists, and can even be uncomfortable working with them.
He cited a scene from the film “Hidden Figures” in which a little-regarded person in a NASA meeting room (Katherine Johnson, played by Taraji P. Henson) was the only person who understood a re-entry equation. “This is what it often feels like to work with an engineering team,” said the AI scientist, who noted that while more people with AI knowledge are joining engineering teams, most companies have hundreds of engineers to only a handful of AI and data scientists.
He also said that non-communication must be overcome, telling the audience about a time when Monolith was working with a large aircraft manufacturer in Europe. “[Our team] was trying to figure out what alpha was in an equation,” he said. “It took us months to discover that it was obviously the angle of attack.”
While most companies have hundreds of engineers, they only have a few data scientists, Ahlfeld said.
In another example cited at the show, Ahlfeld talked about how supplier Kautex-Textron used Monolith’s software to tackle one of its biggest engineering challenges: How to reduce the sloshing noise made by fuel in vehicle tanks. He said the company’s engineers were able to teach Monolith’s software to accurately predict real-world test results largely on their own.
Alin Petcu, Kautex Textron’s global lead on virtual validation, said in a news release that fuel sloshing noise, particularly during deceleration, is “one of the most complicated, multi-physics challenges for our engineering team.” He said using their own engineering knowledge, acoustic data and Monolith’s software, they were able to reliably predict the noise generated by new designs.