Recycling’s Next Frontier: Artificial IntelligenceRecycling’s Next Frontier: Artificial Intelligence
Is AI effective — and affordable — for MRFs and municipalities to use in sorting notoriously problematic food-grade polypropylene?

At a Glance
- Edward Kosior of Nextek and NextLooPP answers questions on AI's role in creating recycling-derived food-grade polymers.
- Once trained, AI is an affordable, technology and is complementary to existing systems, Kosior says.
Accurate material sorting poses one of the biggest challenges to efficient recycling, especially when it comes to post-consumer food-grade recycled PP. In this Part 1 of Packaging Digest’s conversation with Edward Kosior, CEO and Founder of Nextek Ltd and its NextLooPP initiative that aims to create food-grade recycled polymer from advanced mechanical recycling, we learn about how Artificial Intelligence is poised to be the loop-closing changemaker.
In April, Tomra Recycling announced that its GAINnext deep learning-based add-on for Autosort sorting units was capable of performing quick and efficient, large-scale separation of food-grade from non-food-grade plastics for PET, PP, and HDPE.
In this conversation with Packaging Digest, Kosior discusses the potential of this technology and the promise it holds for simplifying the recycling sortation process with cost-effective efficiency.
What’s the difference between Tomra’s GainNext system and other AI sorting systems?
Kosior: Compared with other systems, Tomra’s GainNext leverages decades of AI experience and deep learning to automate complex sorting tasks, such as separating out food-grade packaging. NextLooPP has added our deep knowledge of PP packaging to assist the training of the “deep learning” system, which needs to be followed by decontamination technology to create a food-grade recycled PP resin.
The system stands out among AI sorting technologies due to its advanced deep learning capabilities and integration with multi-sensor systems, the AI is trained on thousands of images, enabling it to classify complex and overlapping objects with high accuracy and it distinguishes materials based on properties such as size, shape, and texture, handling tasks that are challenging for standard optical sorting systems.
How will AI impact packaging UV markers?
Kosior: Since the start of NextLooPP in 2020, one of the main focuses has been to efficiently separate food from non-food packaging. To achieve this, the team successfully trialed UV markers in conjunction with Tomra. At the time, this was the most effective spectroscopic sorting technology to separate the same polymer into food and non-food fractions by adding a fully integrated, coded sorting dimension to the standard NIR/VIS sorting systems.
In the meantime, NextLooPP and Tomra had also been exploring using AI to sort, with NextLooPP supporting the PP field validations conducted by Tomra to test their AI system, GAINnext’s capabilities in industrial conditions.
Given that the AI had to be taught and NextLooPP already had its own highly efficient marker technology plug-and-play ready, the decision was to start with markers and then phase in AI. What we had not expected was the incremental speed at which Tomra grew GAINnext’s capabilities. By early 2024, Tomra had accelerated its GAINnext deep learning technology to separate food and non-food plastics and was using it to identify PP packaging, amongst others. It did not take long to realize this was a real game-changer.
What design suggestions should brand owners consider to boost their packaging recyclability?
Kosior: Instead of a system that relies on labels featuring specific markers, the neural network of the AI system is trained to identify a range of shapes and packaging attributes. Through structured training, it learns to separate out food contact from non-food-contact packaging. As such we need to revise the current design guidelines to take into account how the AI “thinks” to continuously enhance both GAINnext’s and other existing sorting solutions’ capacity.
Certainly, the suggested changes to the packaging will be simpler and more cost-effective than relying on labels and markers, if anything the more stereotypical the packaging, the better.
The principles by which GAINnext recognizes a package are based on object recognition. By segmenting a range of different design factors of the pack the AI gathers the different triggers to build its contextual memory of every pack it is shown.
AI is trained on food package shapes, sizes, dimensions, or other criteria that frequently re-occur. Transparency, opacity, print, shapes, and colors alert the system that is designed to aim for accurate recognition of the sorted PP packaging. The more stereotypical the pack shape, inclusive of readily identified attributes, the higher the rate of identification. Given that PP food trays are predominantly unpigmented or white and rectangular, these are easily picked out.
The likes of ice cream tubs, however, which often are solid white, are likely to be rejected from the food packaging stream as they could be identified as non-food dishwasher capsule packs. This is where design attributes and deeper learning can boost the correct recovery. This brings us back to NextLooPP’s original suggestion of using color or design features to signal whether a pack belongs in the food or non-food category.
How will this use of AI improve efficiency and costs?
Kosior: Using AI to sort means fewer changes required to the packaging design whilst still benefiting from highly efficient sorting at speed (during our latest full-scale trials, Autosort with GAINnext sorted five tons per hour of mixed PP plastic packaging and exceeded 97% food-grade content in the sorted output.)
The technology is a game-changer in terms of delivering food-grade recycled plastics, especially PP due to the much lower costs of implementation, the lower operational costs, and the higher efficiency of the technology. This technology is much faster than hand sorting (at least 60 times faster) without the high operational costs.
Can materials recovery facilities (MRFs) and/or municipalities afford this AI-powered technology as an effective solution?
Kosior: The technology is affordable compared with the other options available to achieve the same results. The big difference is in the capacity to train the system to sort accurately at speed and this will be an additional cost that needs to be applied in the various markets where the systems are used. Overall, this is a complementary technology that works alongside existing systems and will be increasingly competitive. Yet will bring new functionality such as digital passports for packaging that is much less costly than existing systems.
During trials, Tomra’s GAINnext system correctly identified over 97.5% prior food packaging content — an outstanding result that is poised to enable brands to meet the sorting standards required to deliver the food safety authorities’ stringent requirements. The technology is effective because it is based on high-speed image analysis, and once it has been trained to respond accurately it can do so many times more efficiently than hand sorting or other marker technologies.
About the Author
You May Also Like