OPINION: When we started Recycleye five years ago, we encountered significant scepticism around artificial intelligence (AI) as a sorting method in MRFs. Today, almost everyone acknowledges the need for both AI- and NIR-based materials sorting, yet replacing near-infrared (NIR) spectroscopy with AI remains controversial.
AI will never entirely substitute NIR, but it will go much further than many currently believe – so if you’re a diehard NIR fan, this article may be an uncomfortable read.
NIR spectroscopy uses light-based sensing to perform sorts based on material, meaning it is the perfect solution for differentiating PP from PET, which look identical to the human eye.
AI-powered sensing involves a camera which scans items and, using deep learning algorithms trained on a dataset of images, categorises each by material and object, based on visual features.
Consequently, the DMR sorting market can be divided into 3 categories: AI only, NIR only and AI and NIR sorts.
AI only
Object-level sorting is an NIR blind spot, thus, this has been the focus of most early AI applications as there is currently no viable alternative.
For example, at an Urbaser MRF in Spain, our client’s robot extracts high value aluminium beverage cans from lower value aluminium foils and trays, which often contain organic residue, producing a higher quality aluminium output.
Other novel applications include separating PET bottles from trays; black plastics; films from rigids (HDPE vs LDPE); and food-contact packaging from non-food-contact packaging.
Most notably from a regulatory perspective, AI excels in identifying multilayer items – a long-held challenge for NIR, as material-based sorting naturally struggles with composite materials. Simpler Recycling will require liquid carton collection from March 2026, and our AI is already equipping MRFs such as Veolia Southwark, Panda Ireland and Bryson Recycling to extract these items.
Overall, AI is most commonly used to automate quality control at the end of lines, correcting the mistakes of NIR machines upstream.
NIR only
Whilst AI computer vision can see everything the human eye sees, it is also limited to what the human eye sees.
Thus, items which are visually identical cannot be differentiated by AI, yet their difference in chemical properties means NIR can accurately sort them.
For example, separating HDPE Jazz from PP Jazz or PP film from LDPE film are both applications in which AI fails but NIR succeeds.
AI and NIR
This is where 80% of sorting applications exists: where both NIR and AI are capable of the same sort.
In paper and plastics sorting, traditional machinery requires NIR to detect paper and plastics, VIS to detect colours, LASER to detect black plastics and electromagnets to detect metals. Meanwhile, AI sensors can detect all of these classes simultaneously, often to higher accuracies as computer vision is not confused by labels and sleeves.
Whilst AI models are still less experienced in some of these areas, progress is being made quickly, and often AI is outperforming NIR on similar batches.
Beyond detection
Yet performance of a machine is not the only metric that matters to a plant manager. So, how does AI stack up in other departments?
Less maintenance is required on AI systems due to their use of LED lights, compared to the halogens used by NIR systems. Whilst NIR sensors can get de-calibrated by big moisture changes, dust build up or changing temperatures, AI is more robustly unaffected by these factors.
It’s also too expensive to install additional NIR sensors on each outfeed line after an optical sorter to measure the composition of outgoing streams, whilst this can easily be done with lower cost AI sensors. This provides omniscience to the operational team, where previously it’s been difficult to tell how successful the optical sort has been.
Finally, AI’s superior understanding of objects compared to NIR means that it comprehends overlap in more detail, and hence sends more precise calibration information to the nozzle bar, resulting in ejection timings that minimise collateral.
This suggests that AI competes with traditional optical sorters not only on detection capabilities, but also on robustness of hardware.
In conclusion, today, the term ‘optical sorter’ still invariably conjures an image for waste management professionals of an NIR-based sensor and air jet sorter. In the next 5 years, I expect that to change.
Subscribe for free