Learn a predictive model, then reuse it across future computations
Quantum computing has long held transformative promise while remaining inaccessible to most who might use it — a technology of extraordinary potential locked behind extraordinary cost. Researchers from Henan Key Laboratory and Nanyang Technological University have now built classical machine learning models that learn the behavior of real quantum processors and predict their outputs, reducing measurement overhead by more than 99.97 percent. Tested on a 42-qubit superconducting processor, these 'predictive surrogates' suggest that the scarcity defining quantum computing's early era may not be permanent — that knowledge of a machine, carefully learned, might stand in for the machine itself.
- Quantum hardware is so rare and expensive that most researchers who develop quantum algorithms never get to run them on actual machines — the gap between theory and experiment is measured in years and dollars.
- Every computation on a quantum processor demands millions of repeated measurements at frustratingly slow rates, turning even modest research tasks into serious logistical and financial bottlenecks.
- The new surrogates act as digital twins: trained on small datasets from real hardware, they predict future quantum computations on ordinary classical computers, bypassing the queue entirely.
- Crucially, the models come with mathematical guarantees rather than black-box assumptions, and were validated on real tasks — variational optimization and quantum phase identification — without sacrificing accuracy.
- The training data required does not grow proportionally with processor size, opening a credible path to scaling the approach toward the thousands of qubits needed for transformative real-world applications.
Quantum computers process information through the counterintuitive rules of quantum mechanics, promising to solve problems classical machines cannot. Yet they remain rare, temperamental, and prohibitively expensive — most researchers who design quantum algorithms never get to test them on real hardware.
A team at the Henan Key Laboratory of Quantum Information and Cryptography and Nanyang Technological University has found a way around this bottleneck. Published in Nature Communications, their approach trains a classical machine learning model on a small dataset drawn from a real quantum processor, teaching it to predict what the hardware would do — without ever running it again. The result cuts measurement overhead by more than 99.97 percent.
Senior author He-Liang Huang framed the problem directly: quantum processors run at kilohertz rates, and tasks requiring millions of repetitions become crushing practical obstacles. The surrogates replace those hardware runs with fast classical predictions. What distinguishes this work is its rigor — the researchers mathematically characterized the conditions under which the models remain reliable, accounting for noise, input complexity, and processor scale.
Validated on a 42-qubit superconducting processor across two real tasks, the surrogates performed with high accuracy. More importantly, the amount of training data required does not scale up as processors grow larger, suggesting the method could extend to machines with thousands of qubits — the scale at which quantum computing might genuinely transform chemistry, materials science, and physics.
The team plans to deepen the theoretical foundations, expand to other quantum platforms, and broaden the range of tasks the models can handle. The larger ambition is democratization: allowing researchers worldwide to benefit from quantum hardware without owning it or waiting years for access.
Quantum computers have always promised to solve problems that classical machines cannot touch. They process information using the strange rules of quantum mechanics—superposition, entanglement, the whole counterintuitive toolkit. Yet for all their theoretical power, they remain locked behind a wall of cost and scarcity. Only a handful exist worldwide. They are temperamental, expensive to build, and even more expensive to run. Most researchers who dream up quantum algorithms never get to test them on actual hardware.
Now a team at the Henan Key Laboratory of Quantum Information and Cryptography and Nanyang Technological University has found a way to sidestep that bottleneck. They have built what amounts to a digital twin of a quantum processor—a classical machine learning model that learns how a quantum computer behaves, then predicts what it would do without actually running it. The work, published in Nature Communications, cuts the measurement overhead by more than 99.97 percent.
The idea is elegant. Train a classical algorithm on a small dataset pulled from a real quantum processor. Let it learn the relationship between inputs and outputs. Then, for future computations, run the prediction on an ordinary computer instead of queuing up time on expensive hardware. He-Liang Huang, the senior author, explained the appeal plainly: quantum processors are rare, they are slow, and most researchers cannot access them. "A full quantum circuit is often repeated only at kilohertz rates," he said. "When a task requires millions of repetitions, this quickly turns into a serious practical bottleneck." The surrogates solve this by replacing those millions of hardware runs with fast classical predictions.
What makes this different from other machine learning shortcuts is that it comes with mathematical guarantees. The researchers did not build a black box. They rigorously mapped out what could go wrong—noise in the system, the dimension of the inputs, the size of the processor—and proved their model would still work. They tested it on a real superconducting quantum processor with 42 programmable qubits, running it through two practical tasks: optimizing variational quantum eigensolvers and identifying non-equilibrium quantum phases of matter. In both cases, the surrogates achieved high accuracy while slashing measurement costs by more than 99.97 percent.
Perhaps most striking is that the amount of quantum data needed to train the surrogates does not balloon as the processor grows larger. This suggests the approach could scale to machines with thousands of qubits—the kind that might actually solve real problems in chemistry, materials science, or fundamental physics. Huang sees the path forward clearly: "Traditionally, every new computation must be carried out directly and entirely on quantum hardware. Our results show that, for many practically important tasks, it is instead possible to learn a predictive model of the processor and then reuse that model across several future computations."
The team is not stopping here. They plan to deepen the theoretical understanding of why these surrogates work so well, to extend the framework beyond standard qubits to other quantum platforms, and to make the models powerful enough to handle a wider range of real-world tasks. The long-term vision is to democratize quantum computing—to let researchers, engineers, and scientists benefit from quantum hardware without needing their own machine or years of waiting for access. For a field that has always been bottlenecked by scarcity, that would be a genuine shift.
Notable Quotes
A full quantum circuit is often repeated only at kilohertz rates. When a task requires millions of repetitions, this quickly turns into a serious practical bottleneck.— He-Liang Huang, senior author
This work opens a path towards democratizing access to advanced quantum hardware.— He-Liang Huang
The Hearth Conversation Another angle on the story
So the core problem is that quantum computers are expensive and rare. But why not just build more of them?
Because they are extraordinarily difficult to build and maintain. A superconducting quantum processor has to be kept at near absolute zero. The qubits are fragile. Errors creep in constantly. The infrastructure alone costs millions. So instead of solving access by building more hardware, this team solved it by learning to predict what the hardware would do.
And they trained this prediction model on real data from an actual quantum processor?
Yes. They took a small dataset from a 42-qubit machine, fed it to a classical machine learning algorithm, and the algorithm learned the pattern. Once trained, it could predict the outcomes of new quantum computations without ever touching the quantum hardware again.
But machine learning models are often black boxes. How do they know the predictions are trustworthy?
That is the key innovation. They did not just build a model and hope. They proved mathematically what could go wrong and showed the model would still work within those bounds. They mapped out the sources of error—noise, input size, processor size—and proved the predictions would remain accurate.
The 99.97 percent reduction in measurement overhead sounds almost too good to be true.
It is real, but it is also specific to the tasks they tested. The surrogates replace millions of expensive hardware runs with fast classical predictions. But they only work for tasks where you can afford to train the model first. For completely novel problems, you still need the hardware.
What happens next? Do they think this scales?
They believe it does. The amount of training data needed does not grow dramatically as the processor gets bigger. So theoretically, you could train a surrogate on a 42-qubit machine and use it to predict outcomes for machines with thousands of qubits. That is what they want to test next.