Deep Learning: The New Face in the Wood Recycling Industry
At what stage of development is deep learning in the field of recycling at the moment? How has it evolved over the years?
At TOMRA, we believe deep learning will drive material circularity. Other sorting systems separate materials by type, color, or density. Our deep learning system uses object recognition via RGB cameras. Our experts train the network with thousands to millions of images until it can distinguish visual characteristics, such as a material’s shape or a specific feature like a bottle cap.
In 2019, when we launched the industry’s first deep-learning technology, now known as GAINnext™, it could only solve one application: the purification of PE streams by removing silicone cartridges. Over the years, our experts have trained it to tackle increasingly complex sorting tasks—tasks that usually require human intelligence. We have applied this solution to different areas, including paper, aluminum packaging, and wood sorting.
What are the benefits of using deep learning in sorting, and how has TOMRA’s GAINnext™ revolutionized sorting?
Such advanced technologies significantly improve the sorting and classification of recyclable materials and help increase the efficiency and automation of plants. They are poised to increase sorting granularity, which is indispensable for a genuine circular economy. Additionally, these technologies are unlocking completely new applications that were previously unsolvable.
Take our wood sorting applications as an example: In 2022, GAINnext™ as the first solution on the market enabled the sorting of natural wood (wood A) from processed wood (wood B). Today, it is also able to recover MDF and we can clean and sort construction and demolition waste wood. These are very challenging tasks due to the material's identical type, and they were once considered impossible.
Deep learning technology is especially powerful when combined with other sensors. By combining it, for example, with NIR systems or adding a deep learning sorting step to an X-ray sorting step, we achieve even higher sorting accuracy and efficiency, further advancing the capabilities of our sorting processes.
How does TOMRA intend to scale up GAINnext™ over the coming year?
We will continue to expand our GAINnext™ ecosystem in the future, enabling it to solve even more complex sorting tasks in different areas. More than 100 of our solutions have already been installed worldwide, and we’re now seeing more and more customers implementing our deep learning solution on larger-scale processes. Thanks to this internal knowledge and experience, we are very well-positioned to scale up GAINnext™ with increased speed—GAINnext™ has been developed completely in-house by our R&D teams and AI experts. This in-house expertise is our strength.
How will PPWR influence innovations in recycling technologies?
The European Packaging and Packaging Waste Regulation (PPWR) will be a critical focus in 2025 due to its far-reaching impact on the industry. The PPWR includes provisions related to wood recycling, setting specific recycling targets for wood packaging: 25% by 2025 and 30% by 2030. Additionally, the PPWR emphasizes the importance of recyclability and reusability, encouraging the use of sustainable materials like wood.
We expect these targets to drive innovations in eco-design and recycling technologies, including advanced mechanical recycling, as reaching these targets requires consistently high-quality recovered materials. Sorting is a crucial step in this process; it is key to feeding as many materials as possible back into the cycle. We need sophisticated solutions like GAINnext™ to help us achieve the highest possible sorting granularity and we need to act with urgency to reach the targets.
What may prevent companies from adopting the use of these deep learning technologies? What are some pressing concerns or considerations and how may they be addressed?
The adoption of deep learning technologies in the wood recycling industry offers significant potential, but it also comes with challenges that may prevent companies from fully embracing these solutions. One of the primary concerns is the complexity of developing and deploying deep learning models, which requires specialized expertise and a robust infrastructure. Developing reliable neural networks capable of distinguishing materials in real-time is far from a plug-and-play process. It involves continuous training, data collection, and validation by highly qualified professionals in fields such as AI, data science, and material recognition.
For most companies operating in the recycling sector, building this kind of expertise in-house is neither feasible nor cost-effective, which is why partnering with global solutions providers like TOMRA is a more practical approach. We have invested heavily in AI-driven solutions over the past decades, have vast experience in this field, and use this knowledge not only to advise on the best suitable technology but also on a tailor-made process that fits exactly our customers’ needs.
