At PRS Europe, BP&R sits down with Alberto Piovesan at TOMRA Recycling to discuss the importance of neural networks and flake sorting.
TOMRA
[GD] How does it feel like to have won the ‘Recycling Machinery Innovation’ category with GAINnext?
[AP] We are thrilled to have received this industry accolade for our innovative technology, GAINnext. It's a testament to the work we've been doing.
We’re deepening our knowledge of neural networks. Deep learning, a subset of AI, allows a machine to think like a human. We input millions of images into the neural network to train it to detect and recognise objects based on visual characteristics, like specific product shapes or dimensions. We pair this with more traditional sensor technology.
We need to start looking beyond NIR or VIS sensors because we’re collecting a lot of data. How can we process this amount effectively? When properly trained, deep learning systems can unlock new ways of sorting highly complex materials.
[GD] I know one of TOMRA’s most recent advancements is flake sorting. Can you explain what it is?
[AP] In advanced mechanical recycling, it’s impossible to achieve purity levels that are high enough for the extrusion process without a flake sorter.
This technology can detect the different types of polymers (PET, PP and PE, among others), colours and contaminants. With PPWR, there’s a greater focus on removing contaminants such as BPA, the industrial chemical bisphenol A. We can target and detect materials that generate BPA or are responsible for BPA, such as polycarbonate. We can break down the contamination and purify the flakes.
This also makes the extrusion process highly efficient. If you have good pre-sorting and flake sorting processes, the yield can be very high throughout. We provide pre-sorting and flake sorting as a package, and we share the technology, the data and the AI between them. The full solution can balance out any type of input contamination. This unlocks opportunities for recyclers in terms of reduced costs, higher margins and a wider choice of feedstock material.
[GD] How has TOMRA’s view on AI evolved since introducing the industry’s first-ever deep learning technology for waste sorting in 2019?
[AP] 2019 was when we first launched our deep-learning solution, GainNEXT. Back then, it could only remove silicone cartridges from the PE stream. This requires a well-trained neural network because the material is the same and the colour can be any depending on the brand and the design. Sometimes, the objects we target are broken, crushed or have something else overlapping them.
It proved that we could teach something to the neural network, and this could learn and execute the task efficiently. It worked across many locations, from the field to big scale, so we started launching further application packages.
These include PP, PET and HDPE food-grade applications, which enabled food- vs. non-food sorting for the first time. We launched the aluminium UBC cleaner, some complex wood sorting applications and more. A neural network needs to know what a PP food bottle or a PP non-food container look like. This requires a skilled team of AI and industry experts as materials evolve constantly depending on the country, the season, the branding, etc.
We have a great system that can learn. Now, it’s a matter of understanding what to teach it and how to continue to build a solid infrastructure around it.
[GD] What trends in AI and recycling do you expect to see going forward, and what would you like to see more of?
[AP] I’d like to see more recycling in general. PPWR is giving it a nice boost while also bringing some challenges.
What we’re trying to do with our technology is increase the yield and lower the operational costs. We know we have a product that can distinguish any type of material and colour. But how do we keep making it more and more efficient so our customers can increase their margins in a challenging market situation?
At TOMRA, our customers are our priority. As an international company, we receive information from all over the world. We understand how things are done in other places, helping us see what works and what doesn’t. This way, we can consult our customers on a specific issue and discuss advantages and disadvantages to reach a solution. They come to us for the quality of our machines, the design process and our consultative approach. It’s all about extensive knowledge paired with the right technology.
[GD] There still seems to be a lot of scepticism around AI. How would you respond to it?
[AP] AI is a term that is often misunderstood. AI in its most basic definition is just a machine’s decision-making process. Those who claim that it isn’t new, which AI are they referring to?
Today’s neural networks are useful. They can think and act like a human. We’re still trying to grasp the breadth of things they can do. If we can improve the quality of our decision-making, it’s a win-win. We widen the spectrum of what can be sorted, extracting more fractions.
It comes down to educating customers. That’s why I care to make the distinction between AI and the neural networks used for the latest deep learning systems. We need to have faith in developing new technology because it can help us become more efficient. Informing people is the first step to making them realise the possibilities available.