Editor Rebekah Jordan spoke with Indrajeet Prasad, product manager Deep Learning at TOMRA Recycling, to learn how its GAINnext system uses deep learning to identify and separate food-grade plastics for PET, PP & HDPE at high purity. Like all technologies, no matter how advanced, deep learning has its limitations. Prasad delved into the challenges associated with using this technology and shared his vision for its future in effectively recycling food-grade plastics and enhancing sustainability.
Key Highlights:
- TOMRA launched its GAINnext system designed to allow precise separation of food-grade from non-food-grade plastics in materials like PET, PP, and HDPE.
- Using an RGB camera, GAINnext can recognise the types of objects based on shape, size, dimensions and more. When materials are identical, deep learning is capable of identifying hard-to-classify objects, such as detecting whether a bottle is food- or non-food.
- The purity rates that this solution is achieving – upwards of 95% for packaging applications in customers’ plants across UK and Europe – will open up opportunities for new revenue streams for its customers.
TOMRA
Indrajeet Prasad, product manager deep learning, TOMRA Recycling
Can you elaborate on TOMRA's journey with incorporating AI and deep learning into your plastic sorting solutions?
Our R&D teams have been using AI for years to develop intelligent sorting solutions. In fact, the working principle of our very first recycling sorters – some 30 years ago – was based on AI. Even then, our machines were able to make decisions just like human beings and decide which materials to eject and reject.
In 2019, we launched GAIN in an application which used deep learning neural networks to remove silicone cartridges from polyethylene streams, soon followed by an application for wood chip classification in 2022. Since then, our deep learning engineers have trained artificial neural networks with millions of object images to solve some of the world’s most complex sorting tasks.
In March 2024, we rebranded our GAIN technology as GAINnext in tribute to the product’s vast evolution. GAINnext is available as an add-on to our AUTOSORT unit which is TOMRA’s best-selling, powerful, multifunctional sorting machine. We introduced five new applications of our GAINnext technology, three of which are groundbreaking applications to separate food-grade from non-food-grade plastics. Using GAINnext, we can now – for the first time ever – quickly and efficiently separate food-grade from non-food-grade plastics for PET, PP and HDPE on a large scale.
We also launched two non-food applications that complement our GAINnext ecosystem: an application for deinking paper for cleaner paper streams, and a PET cleaner application for even higher purity PET bottle streams.
Today, we have more than 100 AUTOSORT units with GAINnext installed at material recovery facilities across the globe.
How does it work and what are the major benefits of using AI in these methods?
The value of deep learning in sensor-based sorting lies in object recognition. Using an RGB camera, GAINnext can recognise the types of objects based on shape, size, dimensions and more. When materials are identical, for example, deep learning is capable of identifying hard-to-classify objects, such as detecting whether a bottle is food- or non-food.
However, the true power and benefits come from combining deep learning (GAINnext) with traditional sensor-based sensors (AUTOSORT). This combination delivers outstanding results and refines the sorting process to an even higher sorting granularity. This, in turn, brings us closer to material circularity, enabling our customers to upgrade recovered plastics and achieve bottle-to-bottle quality material.
The purity rates that this solution is achieving – upwards of 95% for packaging applications in customers’ plants across UK and Europe – will open up opportunities for new revenue streams for our customers.
How does AI tackle problems such as hygiene concerns and increasingly stringent industry regulations?
Globally, we are seeing increased demand from consumer brands for more high-purity recycled content in the food, cosmetics and pharmaceutical industries. At the same time, there is a regulatory push for boosting plastics recycling and safe recycled content in Europe. The Packaging and Packaging Waste Regulation (PPWR) sets targets for recycling and recycled content, while Plastics Recycling Regulation (EU 2022/1616) dictates what can go into food contact plastics and sets requirements relating to processes and operations to decontaminate plastic during recycling.
These more stringent EU regulations are driving the push towards greater circularity, and meeting these demanding standards is very challenging, with 95% of the recycled content needing to come from former food packaging. Hygiene concerns add a further layer of complexity to handling food waste in recycling. We need to leverage the latest AI technologies to resolve these complexities.
Until now, food-grade sorting has proved a real challenge for the industry as food and non-food packaging are often made of the same material and visually very similar which makes it difficult for any sorting system on the market today to differentiate and separate. However, using deep learning helps to resolve the challenges associated with the complex area of food vs. non-food contact applications because it can quickly and efficiently separate food-grade from non-food-grade plastics for PET, PP and HDPE.
Are there any challenges or potential risks associated with using these intelligent methods?
Deep learning technologies require much more sensor data to train the neural networks than conventional machine learning, so we made sure we adapted to these requirements very early on, investing in AI and creating an IT infrastructure to develop scalable solutions.
There is a lot of hype around AI but it’s important to remember that, as with every technology, AI has its limitations. It is only with in-depth industry knowledge that we can harness the power of all technologies to get closer to true material circularity. And at TOMRA, we believe the greatest value comes from combining traditional sensor-based sorting with the latest deep-learning technologies.
How will this technology help to close the loop on many plastics in food-grade applications and drive sustainability?
Achieving higher-quality recycling is a necessity and it is only through the use of the most advanced deep learning technology that the industry will be able to deliver high-value outputs to meet increased recycled content targets. The transition to the circular economy is resulting in more regulations which can only be achieved by tapping into new technology such as GAINnext to deliver more granular sorting. GAINnext is already creating new opportunities for closing the loop on many plastics in food-grade applications.
Looking ahead, how do you see AI and deep learning evolving within TOMRA's plastic sorting solutions? Are there any specific areas of development you're excited about?
We have used AI technology to improve sorting performance for decades, but our latest application marks another industry first. We are just at the start of the next stage of the resource revolution, and the use of AI will boost material circularity at a time when it is most needed.
AI has the power to transform resource recovery as we know it, and our latest sophisticated applications of deep learning and AI reinforce our position as a pioneer in this field. With its sophisticated use of deep learning, GAINnext enables food-grade sorting and bottle-to-bottle quality, tasks that have posed significant challenges for our industry for many years.
We are already working on further new GAINnext applications which will be released over the course of the year, and we will make this innovative deep-learning technology available to our metals’ customers for the first time.