There is a common misconception that AI is new to the sorting industry, but AI has always been part of Tomra’s DNA. Even the company’s first Autosort systems, dating back some 30 years, used basic AI principles to make intelligent decisions about which materials to keep or reject
The ability to make decisions like a human is the very definition of AI. This type of technology is not just a trend. What has changed is the use and scale of deep learning, a subset of machine learning that allows sorting systems to move beyond the physical limits of traditional sensors.
A POWERFUL EXTENSION, NOT A REPLACEMENT
Deep learning is a specialised branch of machine learning, built on neural networks trained on huge volumes of raw data to recognise intricate patterns and apply them to new material. It is a powerful extension of proven sensor-based sorting, such as near-infrared (NIR) and colour-sensor systems. It adds value through visual classification that conventional sensors cannot achieve alone. Its strength lies in object recognition using full-colour cameras, meaning these systems see what the human eye can see.
The Gainnext ecosystem is a great example, solving previously impossible tasks such as distinguishing between food-grade and non-food-grade PET. These materials can share an identical composition and spectral signature, making them invisible to conventional optical sorting.
The Gainnext PET Cleaner application, combined with Autosort, delivers purity levels exceeding today’s industry standards, eliminating hard-to-detect items such as opaque white packaging, textiles and foils from PET bottle streams. This produces recycled material that closely resembles virgin raw materials, ensuring PET is not merely downcycled but recovered for the most demanding high-end applications, such as food-grade packaging and high-quality textiles.
The same principle now extends to food-grade PET trays, historically inseparable from non-food trays using conventional optical sorting as they share the same material composition and spectral signature. Gainnext overcomes this by distinguishing takeaway or supermarket trays from non-food trays, including consumer goods and medical packaging, achieving purity levels above 95% without manual sorting, according to the company. As demand for food-grade rPET grows, PET trays are becoming a critical new feedstock alongside bottles.
Regulatory pressure is intensifying, and the EU’s Packaging and Packaging Waste Regulation (PPWR) introduces stricter requirements for recyclability and recycled content. By enabling tray-to-tray recycling, Gainnext supports the transition to closed-loop plastic packaging systems, helping recyclers secure feedstock, meet regulatory requirements and remain competitive.
FROM REACTIVE TO PROACTIVE PROCESS OPTIMISATION IN SORTING
AI-powered sorting systems generate data on material composition, sorting efficiency and equipment performance. Intelligent camera systems can also monitor material flows throughout the plant, tracking purity, flagging losses and supporting regulatory compliance.
Historically, operators relied on retrospective, manual analysis to fix today’s problems with yesterday’s data. The industry is now moving toward continuous, data-driven optimisation, where decisions are based on live material composition trends, allowing for far more stable output, the primary requirement for bottle-to-bottle applications. Tomra expects greater integration through systems like Tomra Local Control, which enables operators to manage multiple sorters centrally and react faster to changing conditions on the plant floor. This keeps sorting stable even when infeed material is highly variable, allowing the entire line to be adjusted as one cohesive system. Complementing Tomra Local Control, the PolyPerception Waste Analyser acts as an automated, continuous quality auditor, using cameras to track material composition from infeed to output and provide instant visibility into material loss and purity at strategic points across the sorting circuit.
PolyPerception has introduced an AI-native platform that enables operators to interact with plant data through a conversational AI interface. Similar to using a digital assistant, users can ask questions in natural language, generate bespoke quality reports, investigate process deviations and receive insights tailored to their sorting processes and goals, without the need to manually analyse complex datasets or dashboards.
Users can also explore performance trends and process behaviour by asking questions such as why purity levels have changed, where material losses are occurring or which sorting line is underperforming, receiving immediate, data-backed insights.
The platform also addresses one of the industry’s fastest-growing risks: fire hazards from items such as vapes and batteries. Using a similarity search, operators can flag one problematic object and instantly locate every visually similar item, triggering automated alerts without retraining any models. For compliance, it automates sampling at the infeed and key outputs, replacing manual reporting with documented, real-time evidence for audits.
DATA AS INFRASTRUCTURE FOR THE GREEN TRANSITION
As global regulations tighten, the ability to provide a digital birth certificate for recycled material through data will be just as critical as the physical sorting process itself. This data will help manufacturers prove, for example, that less than 5% of the input material consists of non-food-grade PET.
This is why AI-driven analytics and plant-wide data intelligence will play an increasing role in waste analysis, extending transparency throughout the sorting process, not just at the sorter itself. As both regulatory and consumer pressures build, deep learning offers a powerful solution to advance the circular economy. By unlocking its potential, Tomra aims to create new markets for higher-value products, further stimulating growth and sustainability.
For more information visit: www.tomra.com/waste-metal-recycling
