Metaspectral pairs next-generation sensors with AI to advance recycling

Canada recycles only about nine percent of its plastic waste, with much of it ending up in landfills.

Metaspectral CEO Francis Doumet believes that one of the reasons for this is because we can’t properly separate plastics at recycling plants. “The more we’re able to separate materials, the more we can divert from landfills,” he says, having worked on a plastic recycling project for the past three years.

The company specializes in next-generation infrared sensors that gather 100 times more information than a conventional camera. When paired with AI, they’re able to handle complex data, and large volumes of it.

Through funding from various initiatives—including the CleanBC Plastics Action Fund and Digital—and a partnership with Merlin Plastics and UBC, Metaspectral has installed three sensors in a local recycling plant.

“The sensors themselves aren’t new, but historically it would take an army of PhDs, days and sometimes weeks to analyze the data and come back with the results,” Doumet says. “That’s where we come in—we’ve built a series of technologies and innovated in data compression and AI, enabling us to derive insights from the data in milliseconds.”

The sensors also allow the recycling plant to better identify complex packaging and plastics that look identical to the human eye.

Doumet explains that many consumer products keep innovating in packaging, which makes it cheaper and lighter but also more complex to recycle because of its multi-layering. He gives the example of potato chip bags that have an aluminum coating, and then a middle layer like nylon to keep the oxygen in.

However, with the company’s technology, it’s possible to identify the various layers. Instead of running AI over the entire image a single time, each pixel is a data point to examine how the light reflects back and the different frequencies that are captured. The plants run from 16 to 24 hours a day, and each image can be about a gigabyte in size.

“It’s really hard to run the AI on these systems in real time,” said Doumet. “You can sort materials that are distinct from each other without AI, but if you want to do more complex things like swap between different grades of the same material, that’s when you really need AI.”

Another reason this is needed is to audit recycled material. The current machinery is able to get 90-to-95 percent purity, but Doumet says that customers are asking for at least 98 percent. The sensors are able to identify the grades of material and the purity of the material, which prevents recyclers from being lowballed in terms of price.

The same sensors that can audit the material can be connected to a mechanical sorting device, such as a robotic arm or an air system to blow the plastic piece into one bin or another. While not needed in all plants, Doumet highlights that this could fully automate the process.

But it doesn’t come without its challenges. “When we started, we thought recycling was going to be like any other kind of sorting application but it turns out that recycling is actually one of the more complex sorting tasks,” said Doumet, adding that the majority of the packaging is transparent, causing the sensors to also pick up what’s underneath the objects. 

While Metaspectral has worked to train its AI models on all the different conditions, stacked packages and ones with food residue or dirt from sitting in a blue bin outside can sometimes pose challenges.

“The better we can identify complex packaging, the better we can sort it, and that leads to a higher quality of recycled material,” he says. “And that then leads to more adoption of recycled material in industry.”