For the last four years, Jan Laenen, a senior director of engineering at Kellanova, has had a lofty ambition: to ensure that every crisp, salty chip coming off the production line is, in essence, the perfect Pringle.
No broken dough... no sticky chips caught in the fryer... Just a perfect, thin, saddle-shaped golden sliver. "We want to make a perfect chip... the same crunchiness, the same taste and feel when you open the can," he said.
Now, thanks to a $4 million to $5 million investment, a yearslong partnership with industrial tech company Siemens, and a major AI development project, he's a little closer to that reality.
In that time, Laenen's team has been working with Siemens to build a real-time "digital twin" of the Pringle dough as it moves through the production line of the company's factory in Poland.
New sensors and live machine data were integrated to create live digital versions of the dough, capturing minute variances in raw materials -- like the particle size of the flour. That data now feeds into an AI model that can pre-emptively suggest how to tweak the machines to account for those variances, ensuring every chip is consistent, and flawless.
"Every potato batch you get is going to be different," said Cedrik Neike, member of the Managing Board and CEO Digital Industries at Siemens. Now, "independent of the input, we make sure that the Pringles taste would be exactly what you want it to be."
Digital twins on the factory floor
But the Pringles project is also the latest sign that digital twins, a years-old concept, are starting to scale in manufacturing, boosted by investments from companies like Nvidia. Siemens said it's worked on digital twin projects for everything from rockets in space to batteries to microchips -- besides snack chips.
Today, 705 million pounds of Pringles are produced annually across three main factories in Poland, Belgium and Jackson, Tenn., in the U.S. Late last year, Pringles-maker Kellanova was officially acquired by privately-held snacks company Mars, which now handles U.S. manufacturing of the chips. Timelines for integrating the businesses vary by region, and in Europe the two businesses are still operating side by side.
The digital twin project is live and operational on just one line in Poland, where Laenen says it is delivering a 10% improvement in the quality of Pringles produced, a 13% reduction in waste, and an overall 40%-plus return on the $4 to $5 million the company invested in the project. The system will be expanding to Belgium, with plans for it to arrive in the U.S. in 2027.
"The idea is to scale it to probably all the lines in the end," Laenen said.
The project is emblematic of the way digital twin technology has matured across the sector, said Evan Brown, a principal analyst in Gartner's emerging technologies trends and invest practice. "In the last year, it's really taken off," he said.
When they first started gaining traction five years ago, digital twins were often just 3-D models with relevant data tagged to them. Over time, however, they operate closer to real-time and began offering not just a view of what's happening in a factory, but AI-powered recommendations on how to improve or optimize production, Brown said.
The next step in the evolution, he said, is granting AI autonomy to tweak machine settings and operations based on its own recommendations. In most cases, however -- including that of Pringles -- companies prefer to have a human worker make those calls.
What goes into a Pringle
More than 200 parameters go into creating the "perfect Pringle," from the flour's particle size and the dough's humidity to the geography where the potatoes were harvested, according to Laenen. Even the qualities of potatoes coming from the same supplier in the same location can vary significantly depending on the time of year.
Those differences mean the dough is constantly changing. To ensure consistency in the finished product, machines need to be tweaked and adjusted to account for the small variations.
"A different dough could mean a different texture. It has an impact downstream," said Francesco Ielmini, a senior project manager at Siemens who worked on the collaboration.
In the past, this was all done by look and feel, Laenen said. "Doughmakers" in the factories would grab a piece of dough off the production line and stretch it out, or grab chips just out of the fryer and weigh and measure them. If they detected any aberrations, they'd tweak the machines after the fact.
Now, 200 data points are captured by sensors on the production-line machines every millisecond, and the data is processed inside the factory on a Siemens Industrial Edge Device platform.
All the while, an AI model is running real-time simulations to predict how the Pringles will turn out -- and confirm they meet the standard of the ideal chip. If the model detects anomalies in new raw materials, for instance, it will suggest small adjustments, like tweaking the amount of oil or water going into the recipe.
If the adjustments aren't enough to alleviate the problem, the system will turn to a larger AI model that runs in the cloud and has access to more historical data, to retrain it. Both these AI models use conventional machine-learning rather than generative AI.
The AI model is continuously improving based on feedback from operators, Laenen said, and there are opportunities to include more variables and parameters over time, including different types of potatoes and varieties of corn and rice.
"Today, we want one type of corn, one type of potatoes and one type of rice. We could say in the future, we can handle two or three different types of potatoes and maybe two or three different varieties of rice and corn because the model has learned how to adapt with it," he said. "We want to, of course, use that variability to still make that same perfect chip."
Comments