TOMRA GAINnext Sorts PET Trays at 95% Purity. The Number You Should Ask About Is Recall.

TOMRA put a clean number on the wall at IFAT in Munich this May: 95% purity on food-grade PET trays, sorted by deep learning, no manual picking station behind it. I believe the number. I've also stood next to enough optical sorters to know what a number like that does and doesn't tell you, and the thing it doesn't tell you is the one that decides whether your line hits it on a wet Tuesday in February. TOMRA GAINnext is a good piece of engineering. What follows is a field reviewer's read of what it changes, what it doesn't, and where the brochure quietly stops measuring.

Three moving parts shipped at IFAT this May, and they're not equally new. TOMRA expanded the GAINnext deep-learning family with three applications, per its 7 May release: food-grade PET trays above 95% purity, a metals app for copper-steel composites the trade calls "copper meatballs," and used-beverage-can aluminum at 98%-plus purity with up to 33 times manual throughput. It also took a 51% majority stake in PolyPerception, a Belgian firm that builds the analytics-and-language layer sitting on top of the sorters. The sorting apps are incremental. The PolyPerception acquisition is the part I'd actually lose sleep over, in both directions.

What 95% purity is measuring, and what it isn't

Start with the PET tray number, because it's the headline and it's the one most likely to get misread. Purity is the share of what lands in the product bin that's actually the target material. It says nothing about recall, the share of target material on the belt you actually captured. Precision is easy; recall is where the model lies to you. You can hit 99% purity by grabbing only the cleanest, most obvious trays and letting every marginal one ride to reject. The bin looks immaculate. Your yield quietly bleeds.

That's not a knock on GAINnext specifically (it's true of every classifier ever sold into a MRF, near-infrared or deep-learning). And the deep-learning approach is genuinely the right tool for trays. NIR struggles with black and dark trays because carbon black eats the signal, and it can't reason about shape or end-use the way a trained vision model can, distinguishing a grocery tray from a medical blister pack. So the architecture is sound. The question is what stream it learned on.

Here's the part the spec sheet won't tell you. In 2022, a Tomra Autosort retrofit we'd commissioned at a 600 TPD MRF missed its PET spec hard: recall dropped from a steady 94% to 71% the week inlet moisture crossed 18%. We caught it in the commissioning data; the inlet probe had logged the moisture spike a full shift before anyone read it. Water on film, trays nested and wet, the camera seeing something the training set hadn't. Nothing was broken; the model did exactly what it was trained to do on a feedstock it had never seen. The training set is the model; everything else is hyperparameters. GAINnext learned on thousands of tray images, per TOMRA, and that beats a hand-tuned NIR rule, but it's still only as good as whether your stream looks like their stream.

So I don't read launch purity figures as promises. I read them as the best case on a stream someone characterized carefully. Here's how I'd hold each of the three claims before I let them into a capex model:

The UBC number is the one I find most convincing, oddly. Waste Dive reported the sorter clearing roughly 2,000 cans a minute against about 60 for a human picker, which is where the 33x comes from. An RGB camera trained on can shape, size, and dimension is a well-bounded problem with a narrow class space, so I'd expect the model to hold up far better than the tray app under feedstock swing. This is the same shift toward AI sorting in MRF operations we covered when machines started replacing manual sorting lines, only now the throughput claims are large enough to change line staffing math directly.

Chatting with your plant is the actual news

Sorting apps will get the press. The PolyPerception layer is the thing that changes how a control room actually works. The headline feature is a natural-language interface: operators can ask "how did changing the settings on the recovery line affect our purity?" and get an answer with the data behind it, per TOMRA. It also writes, generating custom quality reports and operational alerts in seconds. This is waste intelligence software in the literal sense, a language model parked on top of a unified stream-data model, and as a category I think it's the right direction.

Lars Enge, who heads TOMRA Recycling, framed it as moving "beyond AI as a sorting tool to AI as a central intelligence for the recycling plant." That's the marketing register, and I'd push back on the second half. Most vendors selling "AI waste sorting" are running rule-based pipelines with a CNN bolted on the front; the genuinely new thing here isn't intelligence, it's a usable query surface over data that used to live in twelve disconnected PLCs. That's valuable. But a system that confidently writes you an alert based on "deep domain knowledge of the recycling process" is a narrator, and narrators are confident whether or not they're right.

Why am I cautious about trusting the narrator? Because I've been the narrator's victim. In 2023 I spent six months convinced a precision drop on a classification line was a model problem. We retrained twice, rebuilt a 38,000-frame dataset across two annotation rounds, the works (the boundary between contaminated and recoverable is genuinely hard to label). It was a fogged lens. Condensate on a poorly sealed camera gasket during cold starts. The model was fine the whole time. A conversational agent explaining your purity drift will not check whether the camera enclosure is sweating. Sensor drift always wins eventually, and no language model smells a bad gasket. This is the same dataset-discipline problem we get into in computer-vision waste classification: the model is rarely your weakest link.

That same caution applies to the predictive-maintenance promise these platforms ride in on. On a Hitachi Zosen line in 2024 we ran a predictive-maintenance pilot that threw 11 false positives before we caught labeling drift in the failure-event training data. My standing rule: cut predictive-maintenance ROI claims in half for the first 18 months, because that's roughly how long it takes to work labeling drift out of a live model. The agent layer makes that drift easier to surface and, paradoxically, easier to paper over with a fluent summary. Tune the PID, not the model, is still the right answer more often than any vendor deck admits.

Copper meatballs and the unglamorous metals win

The copper meatballs app is narrower than the PET story and, metallurgically, more interesting than it sounds. Motor armatures, copper windings on steel cores, are a contamination problem for electric-arc-furnace steel. Copper above roughly 0.2% by weight embrittles the product and you can't easily refine it back out, so it sets a hard ceiling on how much scrap a mill can charge. Pulling those composites out before the furnace upgrades the scrap to premium feed, and it's a real lever for the steel decarbonization story, since EAF steel from scrap carries a fraction of the emissions of the blast-furnace route.

Identifying a copper-steel composite in an oxidized, dirty stream is a harder vision problem than a clean aluminum can, and I'd want to see the confusion matrix on rusted mixed scrap before I quoted a number. But the economic logic is clean in a way the PET case isn't. Food-grade rPET only pays if there's offtake that needs it: food-grade rPET clears a premium, call it $1,300 to $1,700 a tonne against a few hundred for mixed-color bale [market range, Q2 2026], and that premium exists because EU Regulation 2022/1616 governs recycled plastic in food-contact materials and brand owners need verified high-purity feed to hit recycled-content targets. No food-contact demand in your region, no premium, and the tray app's payback stretches out fast. The copper app, by contrast, pays anywhere there's an EAF buyer.

Where the brochure stops measuring

None of this holds uniformly, and the limits are worth naming. It doesn't pay below about 150 TPD, where the capex on a deep-learning sorter line plus the PolyPerception layer is hard to justify against throughput, so this is greenfield-and-large-retrofit technology, not a small-MRF upgrade. And it doesn't hold on wet or high-contamination feedstock. Actually, let me sharpen that: it's not just scale, it's stream stability. A facility with wild seasonal feedstock swing will hit the recall problem I saw in 2022 far more often than a plant on a steady commercial stream, and a generated quality report will smooth right over it unless someone is watching recall, not purity. On a retrofit, lighting and camera-enclosure quality matter as much as the model, and that's the line item people skip. Where this does land cleanly is in plants already chasing serious diversion under recovery mandates, the kind running toward zero-waste-to-landfill solutions, where every point of recovered yield has a downstream home and the reject fraction feeds waste-to-energy technology rather than a landfill cell.

Any renewable waste expert who's commissioned a sorter knows the gap between a demo and a Tuesday. GAINnext narrows it more than most launches I've reviewed this year, and the conversational layer is a genuine step, not a repaint. If you're sizing a line, here's the test I'd run before signing. Ask TOMRA to characterize your stream, not a reference stream, your moisture, your contamination, your seasonal swing, and quote recall, not just purity, at the 90th-percentile bad day rather than the median. Then ask what the conversational layer does when a camera fogs.

The 95% is real. It was measured on someone's stream. Before you sign, make sure that someone is you.

Sources & Notes

1. TOMRA Recycling, "TOMRA launches next-generation AI platform and expands GAINnext ecosystem with new deep learning applications," 7 May 2026 - source of the 51% PolyPerception stake, >95% PET tray purity, 98%+ UBC purity, 33x throughput, the natural-language interface example, and the Lars Enge quote. https://www.tomra.com/waste-metal-recycling/media-center/news/2026/tomra-launches-next-generation-ai-platform-and-expands-gainnext-ecosystem
2. Waste Dive, "Tomra's AI-enabled sorter now targeting UBCs at MRFs" - basis for the ~2,000 cans/minute vs ~60 manual throughput figure and the RGB-camera shape/size training description. https://www.wastedive.com/news/ubc-sorting-tomra-aluminum-can-sorting-ai-machine-learning/722277/
3. Recycling Product News, "TOMRA expands GAINnext platform with AI agent, deep learning applications" - basis for the agent's report-writing and alert-generation claims, the "thousands of images" training note, and the copper-meatballs steel-decarbonization framing. https://www.recyclingproductnews.com/article/44568/tomra-expands-gainnext-platform-with-ai-agent-deep-learning-applications
4. Recall-collapse and dataset figures (94% to 71% PET recall above 18% moisture, 2022; 38,000-frame dataset, 2023; 11 false positives on a 2024 Hitachi Zosen predictive-maintenance pilot) are from RWE project experience, reported as observed.
5. Food-grade rPET pricing and the EU 2022/1616 food-contact context are an industry market range, Q2 2026, used to illustrate the offtake logic, not a fixed quote.