Virtual lab speeds solar recycling breakthrough
Researchers at UNE have developed an AI platform that could make solar panel recycling economically viable at scale - by screening millions of potential chemical solutions in a fraction of the time traditional methods would require.
The platform addresses a looming crisis: how to recycle the 90,000 tonnes of solar panel waste expected annually in Australia by 2030. To date, effective recycling has remained out of reach due to the complex and expensive chemistry required to recover valuable materials.
The polymer problem
A major challenge in solar panel recycling is removing the polymer coating that covers the panel's silicon wafers.
"The critical first step in recycling is removing the polymer coating, and currently, there isn't a cost-effective way to do that," says project leader Professor Amir Karton, UNE Professor of Materials Chemistry. "The polymer is bonded to the silicon wafer; it's not something you can simply peel off."
Finding effective chemical solutions via traditional laboratory testing would be prohibitively slow. "If you wanted to do this in the lab, screening just a few potential solvents would take months of synthesis, testing, and analysis," Professor Karton explains. "Our virtual platform allows us to screen millions of candidates in a much shorter period of time, dramatically accelerating the discovery process."
Quantum chemistry meets machine learning
The challenge is computational. "Screening millions of chemicals with traditional quantum mechanics is computationally impossible. It would take many years on the fastest supercomputers," says Professor Karton.
Enter Dr Kasimir Gregory, Lecturer in Computational, Physical, and Theoretical Chemistry. He's implementing the machine learning simulations that make rapid virtual screening possible.
Dr Gregory uses machine learning potentials - AI systems trained on real quantum chemistry calculations - to replace computationally expensive traditional equations with faster approximations that maintain accuracy. The system can predict how different solvents will interact with the polymer at the molecular level.
"Dr Gregory is the expert in these complex simulations," Professor Karton says. "His expertise in both machine learning and quantum mechanics is the key to making this rapid virtual screening possible."
Dr Gregory's work includes validating the simulations against known experimental results. In testing, the platform successfully identified D-limonene - a green solvent derived from citrus peel - as one of the most effective options among several candidates, matching what researchers had discovered through traditional lab work.
"Successfully predicting a known effective solvent like D-limonene gives us confidence in the predictive accuracy of our computational models," Professor Karton explains.
The simulations run on the UNE’s high-performance computing facility, on GPU infrastructure funded by the university. "Having this high-performance computing infrastructure at UNE is a game-changer," Professor Karton notes. "UNE's investment in this capability, and the support of the research devision is critical for driving this kind of cutting-edge research."
From simulation to reality
Virtual predictions need real-world validation. That's where a $2.7 million Australian Research Council partnership comes in.
UNE has partnered with the University of Wollongong, UNSW, and the University of Newcastle under an ARC LIEF grant to establish the "Self-Driving Automated Molecular Synthesis and Formulation Platform" - a first-of-its-kind technology for Australia, where robots conduct experiments based on AI predictions.
"Our platform funnels millions of potential candidates down to a few thousand of the most promising ones," Professor Karton explains. "These finalists are then passed to the automated lab for validation."
The automated lab is scheduled to become operational in early 2026, with UNE researchers having remote access. "Computational predictions are powerful, but once predictions are made they must be rigorously tested and proven in the real world," Professor Karton says.
Platform ready, industry interested
"What's so exciting is that after several years of development, our AI-driven platform is now fully operational and actively running high-throughput simulations," Professor Karton says.
While solar recycling is the initial focus, the platform's architecture is deliberately versatile. "We are starting with solar panels because it's an urgent environmental challenge that presents a well-defined chemical problem," says Professor Karton. "But if you think about battery recycling, that's a much, much bigger problem."
Battery waste involves more toxic chemicals that pose greater environmental risks as the industry expands rapidly. "The platform is general purpose by design. It can be applied to any complex chemical challenge," he adds, noting potential applications in battery recycling and advanced materials design.
Industry partners are already expressing interest, though agreements are yet to be finalised. Once optimal recycling solutions are identified, they are potentially patentable - with opportunities for both environmental and commercial returns on UNE's investment in this capability.
Understanding the science: Why AI is essential for chemical discovery
The computational challenge
Chemistry at the molecular level is governed by quantum mechanics - the physics of atoms and electrons. In principle, solving quantum mechanical equations can predict exactly how molecules will behave. In practice, these calculations are extraordinarily expensive.
For a small molecule with just 24 atoms (like caffeine), the quantum mechanical calculations involve tracking over 100 electrons moving through an exponentially complex, multi-dimensional mathematical space. As molecules get larger and more complex, the computational cost grows exponentially - making it impossible to screen the millions of chemical combinations needed for breakthrough discoveries.
How machine learning helps
Traditional quantum chemistry calculations solve fundamental physics equations from scratch every time. Machine learning takes a different approach: train an AI on thousands of examples of real chemical reactions, then use it to predict new ones without solving the full equations.
The AI learns patterns - similar to how a person might recognise that certain types of solvents tend to work well with certain types of polymers, but at a molecular level and with quantum mechanical accuracy.
The speed advantage
For UNE's solar panel project, this means the difference between:
Traditional approach: thousands of years of computing time to screen all possibilities
AI approach: days or weeks to narrow numerous candidates down to the most promising hundreds
The AI doesn't replace laboratory experiments - it makes them practical by identifying which experiments are worth doing.