Waste Not, Want Not: Building Digital Twins to Revolutionize Recycling
How Machine Learning and Near-Infrared Monitoring Could Transform the Way We Sort Our Trash
If you think sorting waste is just a matter of placing plastic in the right bin, think again. Beneath the apparent simplicity of modern recycling lies a highly complex, industrial-scale challenge. Every day, waste sorting plants across the world must make split-second decisions to separate valuable recyclable materials from everything else. But these decisions are often made by machines that haven’t kept pace with modern materials — or with the mountains of waste we produce.
The paper “Towards digital twins of waste sorting plants”, published in Resources, Conservation & Recycling in 2024, introduces a ground-breaking approach to this problem: building “digital twins” of industrial waste sorting facilities using machine learning and near-infrared (NIR) sensor data. The work is not only technically rigorous — it’s inspiring, because it tackles one of the most critical bottlenecks in the circular economy: how to recover more usable material from the waste we generate.
Why This Study Matters More Than Ever
Consider this: the extraction and processing of natural resources accounts for half of all global greenhouse gas emissions and over 90% of biodiversity and water stress. And yet, in the European Union — a global leader in sustainability — only 12.7% of raw material demand is met through recycled materials. If we want to move toward a truly circular economy, we must get better at sorting, and thus recovering, post-consumer materials.
At the heart of this effort are sorting plants, which separate mixed waste streams into fractions that can be reused. But these facilities are hampered by outdated control systems, inflexible machine settings, and, most importantly, a lack of reliable data about how sorting performance changes with the type and amount of waste. This leads to both inefficiencies and material loss. What if we could build an accurate virtual replica of a sorting plant — a “digital twin” — that could learn, adapt, and optimize waste processing in real time?
Digital Twins Meet Industrial Waste: A Bold New Approach
In this remarkable study, a research team from RWTH Aachen University and STADLER Anlagenbau GmbH (Germany) presents a method for creating data-driven process models that form the building blocks of digital twins. Their case study centers on a sensor-based sorting (SBS) unit designed to extract polyethylene terephthalate (PET) bottles from a plastic waste stream. This is one of the most common and crucial tasks in modern recycling.
What sets their approach apart is the fusion of machine learning with NIR-based process monitoring — two technologies rarely combined at this industrial scale. Instead of relying on expert-tuned parameters or oversimplified models, the researchers collected continuous, real-time data using hyperspectral imaging sensors and trained artificial intelligence models to understand how variations in waste composition and throughput affect sorting performance.
How the System Works: From Conveyor Belts to Code
The experimental setup featured a full-scale SBS unit, operating in a closed-loop configuration. Waste materials were presented to the sorting machine via a conveyor belt, separated into eject and drop fractions, then recombined and recirculated to allow extensive testing across many conditions.
Near-infrared hyperspectral sensors — positioned before and after the sorter — captured the chemical “fingerprints” of the materials on the belt, classifying them into seven categories: PET, PP, HDPE, PS, beverage cartons, paper/cardboard, and undefined. Crucially, the researchers processed this rich sensor data to measure “transfer coefficients” (TCs) — that is, how much of each material ended up in the right place.
To model this, they trained three types of machine learning algorithms: polynomial regression, random forests, and artificial neural networks (ANNs). The goal was to predict the TC of each material based on factors like occupation density (how much waste is on the belt) and the share of target material (PET). The best-performing ANN model achieved a mean absolute error of just 3.0% — an impressive level of accuracy given the chaotic nature of real-world waste streams.
A Team Effort: Combining Engineering, Data Science, and Industry
This was no solo project. The research involved a multidisciplinary team of eight experts:
Nils Kroell, Abtin Maghmoumi, Tabea Scherling, Alexander Feil, and Kathrin Greiff from the Department of Anthropogenic Material Cycles at RWTH Aachen University, Germany
Tobias Dietl, Xiaozheng Chen, and Bastian Küppers from STADLER Anlagenbau GmbH in Altshausen, Germany
Together, they integrated expertise in material science, industrial engineering, environmental systems, and artificial intelligence — a testament to the complexity and collaborative nature of sustainability research.
Real-World Results: Simulations, Predictions, and Energy Insights
The study went beyond theory by conducting over 33 hours of continuous experiments and simulating entire SBS cascades (multi-step sorting processes involving roughers, cleaners, and scavengers). Using the trained ANN models, they predicted how changes in throughput and material composition affected purity and yield — two key indicators in recycling performance.
Here’s where it gets particularly impressive:
A 1% increase in belt occupation density reduced the F1-score (a balance between purity and yield) by 0.22%.
A 1% increase in PET content improved the F1-score by 0.19%.
Sorting performance varied significantly between materials: PP, for example, sorted more cleanly than HDPE, which is often entangled with PET bottle caps.
The model even revealed how subtle changes — like fluctuations in the input flow — did or did not influence sorting performance, offering predictive insights that could inform real-time machine adjustments.
In one final demonstration, they simulated three alternative SBS plant designs and showed how varying the arrangement of rougher and cleaner units could improve output purity by up to 4%, while balancing energy consumption. With such simulation tools, plant engineers could optimize both economics and sustainability before ever laying a brick or installing a machine.
Why This Work Is a Leap Forward
This study marks a significant step in bringing digital twins — a concept popularized in aerospace and manufacturing — into the domain of environmental engineering. Unlike past attempts that relied on lab-scale models or physics-based simulations with unrealistic assumptions, this work uses real waste, real machines, and real industrial data. It represents a path from offline estimations to responsive, self-optimizing recycling infrastructure.
Tomorrow’s Trash Will Be Today’s Resource
Imagine a future where every sorting plant has a virtual counterpart — a continuously updated mirror that learns from every item of waste passing through. These digital twins could adapt in real time to seasonal changes, new packaging designs, or unexpected waste flows. They could simulate upgrades before implementation, reduce energy use, and maximize resource recovery with surgical precision.
In such a world, landfills would shrink, emissions would drop, and materials that once seemed worthless would reenter the economy as valuable resources.
Thanks to the work of this pioneering team, that future is a little closer.
Open Science Note: The authors published this study under an open-access license. Their approach — combining hyperspectral imaging, machine learning, and open-source software tools like NumPy, pandas, and scikit-learn — provides a foundation for replication and improvement by the global research community. The work reflects a commendable commitment to transparency, reproducibility, and impact.
This blog post is based on this 2024 Resources, Conservation & Recycling Paper.
If you liked this blog post, I recommend having a look at our free deep learning resources or my YouTube Channel.
Text and images of this article are licensed under Creative Commons License 4.0 Attribution. Feel free to reuse and share any part of this work.