Latest / The Manufacturing Podcast with Fexingo: Factories, Supply Chains, and Industrial Business / How a Danish Factory Uses AI to Sort Plastic Waste
Transcript
- Lucas: Luna, I want to take you to a factory in Denmark — a plastic sorting facility outside Copenhagen — where the sorting line is basically a giant computer vision system. Luna: I've seen sorting plants. They use near-infrared sensors, right? They can tell PET from HDPE pretty well. Lucas: They can, but the problem has always been black plastic — carbon black pigment absorbs infrared light, so conventional sensors can't see it. This facility uses hyperspectral cameras combined with a deep learning model trained on over 10 million images. They can identify 12 different polymer types in milliseconds, including black plastic. Luna: So the AI is doing what? Classifying each individual piece as it flies down the conveyor? Lucas: Exactly. Each item gets photographed, classified, and then a jet of air pushes it into the correct chute. The line runs at about 3 metres per second. The whole process happens in less than 50 milliseconds per item. And their sorting purity is 95 percent — that's way above the industry average of maybe 80 percent for conventional optical sorters. Luna: That purity number is the one that matters economically, isn't it? Because recycled plastic that's 95 percent pure can be sold at a much higher price than mixed plastics. Lucas: It's the difference between a commodity and a specialty material. The factory processes 50,000 tonnes of waste per year. They sell the sorted fractions — food-grade PET, clear HDPE, natural PP — back to packaging manufacturers at prices that compete with virgin resin. The CEO told me the plant is profitable without any subsidy. That's rare in recycling. Luna: Without subsidy — that's the headline. Most recycling operations depend on government support or landfill taxes to make the numbers work. Lucas: Right. And the reason is the AI lets them extract value from streams that used to be landfilled. Black plastic alone — think takeaway containers, some cosmetic packaging — that used to be unrecyclable. Now it's a revenue stream. Luna: I've heard that the European Union's extended producer responsibility laws are pushing more of the cost of recycling onto the companies that make the packaging. Does that help the economics here? Lucas: Hugely. Starting next year, packaging producers in the EU will have to pay fees based on how recyclable their packaging is. If your black plastic can be sorted, your fee drops. So brands are suddenly very interested in sending their waste to a facility that can actually separate it. The Danish plant is seeing contracts grow 30 percent year-over-year. Luna: That's a nice feedback loop: regulation creates demand, technology makes it possible, and the economics improve for everyone. Lucas: If today's episode is useful to you — maybe you're thinking about recycling in your own supply chain, or wondering how AI is changing manufacturing — the way we keep these conversations ad-free is through listener support. It's buy me a coffee dot com slash fexingo. Luna: Yeah, it's a small gesture that really helps us keep digging into these specific factory stories. No pressure at all. Lucas: Back to the Danish plant. One of the more surprising things they showed me was how they handle colour sorting. Conventional systems just separate by polymer type, but this AI also grades the colour — clear, light blue, dark blue, green, opaque. That adds another layer of value. Luna: So a clear PET bottle can fetch a higher price than a green one, even though they're both PET? Lucas: Exactly. Clear PET can be recycled back into food-grade bottles. Green PET usually gets downcycled into fibre or strapping. So by sorting by colour within the polymer stream, they maximise the value of each fraction. Their system can separate 20 different colour grades across the 12 polymers. Luna: That's a lot of chutes. How do they physically manage that many output streams? Lucas: They use a series of cascading sorters. The first pass separates polymer type, then the second pass sorts by colour within each polymer. Each sorter has about 60 nozzles per metre, and they fire individually based on the AI's decision. It's like a printhead, but for plastic bottles. Luna: The training data — those 10 million images — where do they come from? Is it just photos of bottles on a conveyor? Lucas: They built a custom rig that captures images of every item that goes through their pilot line — under controlled lighting, with the same hyperspectral cameras. Then they label each image manually in the beginning, and use a semi-supervised learning loop to improve over time. As the plant runs, the model gets better at edge cases: crushed bottles, wet labels, partially shredded material. Luna: So it's a continuous improvement cycle. The AI learns from its own mistakes in production. Lucas: Yes. And they publish their error logs — they're transparent about misclassifications. That's unusual in the AI space, but it builds trust with their customers, who are basically betting that the recycled material is consistent. Luna: I want to ask about the energy cost. Hyperspectral cameras and deep learning inference require a fair bit of computing power. Does that eat into the margins? Lucas: They measured it. The entire AI system — cameras, computers, inference hardware — draws about 15 kilowatts. For a plant processing 50,000 tonnes a year, that's negligible compared to the conveyor motors and the air compressors for the sorting nozzles. The energy cost per tonne sorted is about 30 euro cents. Luna: That's almost nothing. And it's offset by the higher sale price of the sorted plastic. Lucas: Exactly. The plant's revenue per tonne is about double what a conventional sorting plant gets. Because they sell clean, colour-sorted fractions to packaging companies that used to buy only virgin material. One of their biggest customers is a Danish dairy cooperative that uses recycled PET for milk bottles. Luna: Milk bottles — that's a closed loop. The same bottle goes out, gets collected, sorted, and comes back as a new bottle. Lucas: They claim a bottle can go through that loop up to seven times before the polymer degrades enough that it has to be downcycled. That's partly because their sorting is so pure — they remove contaminants that accelerate degradation. Luna: So the AI isn't just about economics; it also improves the quality of the recycled material, which extends its useful life. Lucas: Yes. And that's the piece that gets lost in a lot of the recycling debate. People argue that plastic recycling is a myth because the material degrades. But high-quality sorting can mitigate that. The Danes are proving it at commercial scale. Luna: What's the biggest barrier to scaling this model? Is it the cost of the AI system, or the availability of training data, or something else? Lucas: The cost of the hardware has come down a lot. The hyperspectral cameras that cost $100,000 five years ago are now about $20,000. The bigger barrier is the waste stream itself. To make the economics work, you need a consistent, high-volume supply of post-consumer plastic. That means a well-functioning collection system. In Denmark, they have deposit-return schemes for bottles and curbside collection. In many other countries, the collection infrastructure isn't there. Luna: So the technology is ready, but the upstream logistics are the bottleneck. Lucas: Exactly. The CEO told me they could build another plant tomorrow if they could guarantee the feedstock. But they're limited by how much sorted plastic they can source within a reasonable transport radius. Plastic is bulky and cheap — you don't want to ship it 500 kilometres. Luna: That's a classic chicken and egg problem. More sorting plants would create demand for collection, but collection needs to improve before plants are viable. Lucas: Yeah. But the extended producer responsibility laws are starting to push that. As brands pay higher fees for non-recyclable packaging, they start funding collection infrastructure. In France and Germany, we're already seeing that happen. The Danish plant is a proof point that if you get the feedstock, the AI can make recycling work. Luna: One thing I'm curious about: how do they handle flexible plastics — like films and pouches? Those are notoriously hard to sort. Lucas: Right now, they mostly focus on rigid plastics — bottles, tubs, trays. Films are harder because they're lightweight and tend to wrap around conveyor rollers. But they're piloting a separate line for films using a different camera configuration and a robotic arm that picks them off. Early results look promising. Luna: So the model is extensible. It sounds like they're building a platform that can be adapted to different waste streams. Lucas: That's the vision. They've already licensed the software to two other plants in Sweden and Germany. Each plant retrains the model on its local waste composition — because the mix of plastics in Sweden is different from Denmark. But the core architecture is the same. Luna: It's rare to see a manufacturing story where the technology, the economics, and the regulatory environment all align at the same time. This feels like one of those moments. Lucas: It does. And I think the key takeaway is that AI is not just about replacing humans on assembly lines. In this case, it's about making a material stream valuable that was previously a liability. That's a very different kind of productivity gain. Luna: Yeah. It's not about doing the same thing cheaper. It's about doing something that was impossible before. Lucas: Exactly. And for listeners who track where manufacturing is heading, that's the signal worth paying attention to: when AI turns a waste stream into a revenue stream, the business case for circularity starts to write itself.