In a study published in Science Advances, the UCL team demonstrated how a quantum processor can improve the reliability of machine learning models used to predict the behavior of complex physical systems. The kind of chaotic, turbulent systems that have historically defeated conventional AI. But the researchers believe the broader story is not really about AI at all. It is about what quantum computing can contribute today, right now, as part of real computational workflows.

“We’re making the predictions that are used conventionally more reliable,” says Professor Peter Coveney of UCL. “It’s not replacing something by something else.”

 

A different path forward

The quantum community has long debated whether current noisy, error-prone hardware known as NISQ devices can deliver meaningful results before the fault-tolerant era arrives. The UCL team has a clear view: waiting is the wrong strategy.

“We’re already showing a significant advantage with a NISQ device,” says Coveney, “and we see a kind of continuous pathway from there into the fault-tolerant era. It isn’t going to mean we have to wait before we gain those benefits.”

Their method, Quantum-Informed Machine Learning (QIML), reflects this philosophy in how it is designed. Rather than routing data back and forth between quantum and classical systems — which is slow, error-prone, and expensive in terms of compute time — the quantum processor is used exactly once, offline, at the start of the process. It learns the key statistical patterns of the data: the underlying physical structure that a chaotic system must obey over time. This compressed summary, which the team calls a Q-Prior, is then incorporated into the training of a classical AI model.

The result is an AI that does not just learn from data, it learns physics. And because the quantum step is a one-time investment rather than an ongoing dependency, the approach sidesteps many of the practical limitations of current hardware.

“We only use this quantum device as a one-time, offline training,” explains first author Maida Wang. “Which means we don’t need to communicate between the quantum computer and the high-performance computer every time. That is very efficient, and not costly.”

 

Making AI more reliable

The systems the UCL team is working on are among the hardest to predict: turbulent fluid flows, blood flow in the human body, atmospheric dynamics. These are chaotic systems, where small differences in starting conditions cascade into wildly different outcomes. Conventional AI models, trained on data alone, tend to drift, making plausible-looking predictions that quietly violate the laws of physics.

The quantum-informed approach keeps the AI grounded. In the published study, the method improved predictive accuracy by around a fifth compared to classical AI and remained stable over the long term. In one of the three test cases — a simulation of turbulent channel flow — the classical AI failed entirely without the Q-Prior. With it, predictions outperformed even the most advanced classical alternatives.

The applications the team has in mind are concrete. “In biomedical research, one goal is to predict blood flow dynamics fast enough to inform clinical decisions in real time. Currently, even running the most accurate simulation on the fastest supercomputer available can take up to 24 hours. When a stroke is happening, that is not a useful timescale. Our team is working toward predictions that can be delivered quickly and reliably enough to support clinicians in making decisions,” says joint first author Dr. Xiao Xue.

In weather forecasting, the team is already applying the method to benchmark data and seeing early improvements in prediction accuracy and physical stability. Here too, the goal is not to replace existing supercomputing infrastructure, but to augment it using a relatively small number of qubits to give the AI a better map of the underlying physics.

“For weather forecasting, we may only need 30 to 50 high-fidelity qubits,” says Wang. “We are not talking about thousands.”

 

Working with IQM hardware

Turning this research into published results also required access to real hardware, and a working relationship with the people who build it.

Throughout this work, the UCL team has run exclusively on IQM quantum systems. Access to quantum hardware at research scale is not straightforward. As Coveney puts it: “In most of the world that I interact with, people saying they’re doing quantum computing don’t ever touch a quantum device. It’s not easy to get access to them at the scales we’re talking about. And often, if you do want to do that, you’d be paying a lot of money.”

The team worked across two IQM chip architectures: the Sirius and the Garnet. Early in the collaboration, when the team was still characterizing what the hardware could do, they ran their methods on both systems to understand the differences. Over time, it became clear that the Garnet chip offered the fidelity their experiments required. “We find out that the Garnet chip definitely has better fidelity and is suitable for this case,” says Wang. The Science Advances paper was ultimately built on results from the 20-qubit IQM Garnet.

That kind of iterative process — trying, learning, adapting — is only possible with a hardware partner who makes it possible. When the IQM device at the Leibniz Supercomputing Centre in Munich encountered problems at certain points during the work, the team was able to shift to IQM’s cloud service, and continue without interruption.

Beyond access, the researchers highlight something less tangible but equally important: IQM’s willingness to treat the collaboration as a genuine exchange.

“Of all the device manufacturers we interact with, you are the most interesting to us,” says Coveney. “You work with us and you genuinely want to know what our views are of the devices. With many other vendors, they just expect you to get on with what they’ve given you. They aren’t really interested in your opinions.”

For researchers pushing at the edges of what current quantum hardware can do, that kind of responsiveness matters. The team used several error mitigation methods in the published work and the ability to work through these questions with a hardware partner who is paying attention, rather than simply managing a queue, is part of what made the science possible.

“There is the idea of a co-design scenario,” says Coveney. “I think that’s really good and important with this technology, because it’s in such a nascent, early state.”

 

The hybrid future

A theme running through all of this work is that quantum computing’s near-term value will not come from replacing classical computing, but from working alongside it.

The UCL team’s vision for what this looks like at scale is clear: quantum processing units bolted onto supercomputers, the way GPUs were integrated into high-performance computing infrastructure. Not a monolithic device sitting in isolation, but a specialist component within a broader computational stack doing the things quantum hardware does well, while classical systems handle everything else.

“I don’t think there’s ever going to be a time when a quantum computer is doing everything,” says Coveney. “There’ll be other things that conventional computers do better, and in general they need to be able to exchange data and interoperate with each other.”

Since this work was published, the team has gone further. A follow-up pre-print establishes the formal theoretical foundations of QIML, with a proven quantum-classical separation at its core. It demonstrates the method on weather forecasting, where predictions stay reliable for longer at ranges that matter most for critical operations like severe weather warnings and energy grid management.

We can clearly see that practical progress does not require waiting for a generation of hardware that does not yet exist. The tools available today, used thoughtfully and in close collaboration with the people building them, are already capable of contributing to problems that matter.