UCL: Hybrid of Quantum Computing and AI Predicts Chaotic Systems More Accurately

A team at University College London (UCL) has shown that a quantum computer can already deliver practical value, not in distant theory but in real forecasting problems for complex systems. Their hybrid scheme combines quantum computing and machine learning so that an AI model predicts the behavior of chaotic processes more accurately over long time horizons. In tests, the approach was noticeably more stable than conventional models, improved accuracy by up to roughly 20% and required hundreds of times less memory.

The systems in question are especially hard to model with classical methods: turbulent flows, the motion of liquids and gases, and processes where a small error quickly grows and wrecks the forecast. This is a typical problem for climate models, aerodynamics, energy and biomedicine. A full numerical simulation of such processes can take weeks on supercomputers, while purely neural-network models run faster but often lose stability when the forecast has to be stretched over time. The UCL team tried a middle path: keep a classical AI model, but use a quantum processor during training to give it hints about the structure of the data.

Technically, the scheme works like this: first, a quantum computer processes the training data and extracts invariant statistical properties, that is, hidden patterns that persist over time even in a chaotic environment. These quantum-extracted features are then used to train an ordinary autoregressive model on a classical supercomputer. The authors call the approach quantum-informed machine learning.

An important point is that the quantum block does not take part in every forecasting step and does not require constant data exchange with the classical side. This lowers hardware requirements and helps sidestep typical limits of today's quantum systems, including noise, errors and measurement instability.

The method was tested on several problems: the Kuramoto-Sivashinsky equation, two-dimensional Kolmogorov flow and three-dimensional turbulent channel flow, which is closer to real engineering conditions. According to the paper in Science Advances, the new scheme improved the accuracy of distribution forecasts by up to 17.25% and preserved the system's spectral structure better, in places gaining up to 29.36% over classical baseline models. For the most realistic scenario, the researchers used a 20-qubit IQM quantum computer connected to computing resources at the Leibniz Supercomputing Centre in Germany. The authors separately note that without the quantum prior, predictions became unstable, whereas with it the model produced physically consistent long-term forecasts and in some cases outperformed leading numerical solvers of differential equations.

Efficiency is a separate question. Talk of quantum computing usually runs quickly into the objection that the gain is too expensive or too fragile for practice. Here the researchers present a more down-to-earth and useful picture: the quantum part does not replace the whole pipeline but compresses complex dynamics into a compact representation. The paper reports a memory advantage of orders of magnitude: data of several megabytes was reduced to a quantum representation on the scale of kilobytes. For scientific modeling this is critical, because memory and bandwidth are often no less of a bottleneck than raw compute.

If the approach can be scaled to larger datasets and real-world observations, the applications are many. In climate science, it could mean more stable models of the atmosphere and oceans. In energy, more accurate design of wind turbines and systems that deal with turbulent flows. In medicine, better modeling of blood flow and molecular interactions. In transport and industry, faster aerodynamics and fluid-system calculations without a mandatory increase in memory costs.

The main takeaway is not that quantum computers are suddenly ready to replace classical supercomputers. Rather the opposite: the study shows a realistic scenario in which even today's limited quantum hardware can strengthen existing AI models in narrow but very important scientific tasks. It is one of the most convincing examples that practical quantum advantage may arrive not by overturning current computing, but through targeted integration into AI pipelines that already work.