Better Forecasts and more detailed Models

In modelling dynamical chaos as well as simulating biological molecules, you and your team at UCL now combine quantum computing with supercomputing or artificial intelligence methods – what does this bring?
Prof. Peter Coveney: First of all, my team and I at the Centre for Computational Science at UCL, are researching how we can advance science using computers. So we have some years of experience working with quantum devices within the Noisy Intermediate Scale Quantum- or the NISQ era. This has taught us to seek every feasible way of extracting a signal from the noise, as it were, of these new kinds of computers. We do that by exploiting symmetries and pseudo-symmetries to reduce the number of qubits required for any given calculation; and we have developed ways of reducing noise and errors using a panoply of techniques. Some of these are taken from the literature, others we have invented.

You have simulated a G-protein-coupled receptor or GPCR. Until now, these very large molecules with more than one million atoms have been difficult to model. How did you approach this this time?
Coveney: GPCRs are large, membrane bound and very important proteins which carry out a variety of signal processing functions in our bodies and, as such, are targets for a variety of drugs. Our approach is based on the Nobel prize-winning method known as “QM/MM”, namely a form of hybrid or coupled, multiscale simulation linking a quantum mechanical region – the „QM“ – with one described using classical physics, called the molecular mechanical or „MM“ region. Most of the atoms reside in the classical region and we use classical molecular dynamics to describe their behaviour. The QM region is where the important biochemical processes take place, in this case known as autoproteolysis, the snipping in two of the protein chain by water molecules. The novel feature in our work  is that we offload the “inner core” of the QM region — which is modelled using a rather cheap and relatively inaccurate method known as density functional theory — onto a quantum computer which is used to compute the detailed electronic state of the set of atoms – which is around only eight of them – using a more accurate wavefunction theory. After sampling the wavefunction on the quantum computer, the wavefunction is brought back onto the conventional machine where its electronic energy is calculated as a function of the "reaction coordinate" as the biochemical reaction proceeds. 

What technical resources did you use for this project?
Coveney: In this project, we were able to use Euro-Q-Exa, a quantum computer based on superconducting circuits with 54 qubits from IQM, which is connected to the Bavarian Energy Aware Software Testbed- oder BEAST system, as well as resources from BayernKI. We used various pieces of software including GROMACS for MD and several pieces of quantum mechanical and QM/MM software, including some new QSCI (quantum selected configuration interaction) software which runs on the quantum device. The entire coupled environment was wrapped in NVIDIA’s CUDA-Q which seamlessly integrates the classical and quantum computers together, and permits different backend quantum devices to be connected.

What do the results show?
Coveney: The results to date show that using the quantum computer produces results which are very close to those obtained without its use, but in fact the models we have used are in sum larger than ones which have been studied using conventional computers alone. So the approach allows us to use quantum devices to look at real and complicated macromolecular systems, even if only a small part of them is being calculated on the quantum device. Our scientific research into this system is ongoing at this time — and it is greatly facilitated by the pipeline we have constructed at LRZ that integrates conventional and quantum computers. 

Quantum computers are still somewhat limited – what constraints did you encounter in hybrid simulations?
Coveney: Yes, they certainly are. Let’s make no mistake about that.  That is why we had to minimise the size of the quantum mechanical region which was offloaded onto the IQM device. Only around six until eight atoms could be studied that way even when employing sophisticated noise and error reduction methods; even so, we produced sampled wavefunctions comprising over one million electronic configurations and these required up to 1,200 GPUs in order to determine the ground state energies through post processing of the data which could then be done in around one hour. 

You now demonstrate that quantum computing can improve artificial intelligence – or more precisely, machine learning: how?
Coveney: Classical machine-learning models for chaotic systems gradually lose track of the invariant statistical properties of the dynamics over long autoregressive rollouts, and their predictions eventually drift or collapse to an unphysical static state. Coupling them with quantum computers can counteract this behaviour, leading to long-term predictions that remain physically consistent with the underlying dynamics.

You named this new computational method Quantum Informed Machine Learning or, for short, QIML. How exactly does it work?
Coveney: QIML has two stages. In the first stage, a sample-based quantum generator on a small register is trained offline on a quantum processor to reproduce the invariant velocity distribution of the system. The result is the Q-Prior, a compact representation of the system's long-term statistical structure. In the second stage, the Q-Prior is injected into a classical Koopman autoencoder as a distributional loss, guiding the rollout to remain statistically faithful. The quantum part is light and trained only once; the classical predictor does the heavy lifting on a GPU. The two are coupled only through the statistical constraint, so no quantum data transfer is needed at inference time.

In which research or industrial fields could QIML become important? Who should take a closer look at your method, and why?
Coveney: In principle, QIML has potential for the prediction of any complex dataset whose underlying structure is difficult for classical machine learning to capture in full. The quantum generator extracts statistical information that purely classical models tend to miss, and passes it to a downstream classical predictor in a form the predictor can readily consume. So far we have validated this on canonical chaotic benchmarks of increasing difficulty (the Kuramoto–Sivashinsky equation, two-dimensional Kolmogorov flow, and three-dimensional turbulent channel flow), reported in Science Advances. The same mechanism is expected to be useful for any field whose data carry rich, non-Gaussian or intermittent structure where classical density models tend to struggle.

You even speak of a quantum advantage – why?
Coveney: The advantage we demonstrate is a representational and memory advantage that translates directly into a forecasting gain. A 10-qubit Q-Prior encodes statistical correlations that a comparable classical density model would need orders of magnitude more parameters to capture, and the resulting forecasts surpass leading neural operators in the most demanding case. A full account of how we frame and bound this advantage is given in the Discussion section of our Science Advances paper.

What technical requirements are necessary for this? For example, does QIML depend on a certain number of available qubits?
Coveney: QIML is deliberately modest in its quantum demands. Our demonstrations use ten to fifteen qubits, with the Q-Prior implemented as a shallow circuit that sits comfortably inside the noisy intermediate-scale regime. What matters more than the qubit count is measurement precision and two-qubit gate fidelity: a small, high-fidelity register is more useful for QIML than a larger but noisier one. The classical side is GPU work, which is why the natural setting for QIML is a hybrid QPU + GPU machine of the kind LRZ is now assembling.

You conducted the calculations on superconducting technology – would QIML also work with other quantum technologies?
Coveney: For the QIML project, we used the LRZ’s HPC infrastructure for the classical part, while the quantum part ran on superconducting hardware from IQM, which we accessed via the IQM Resonance cloud platform. The method itself is architecture-agnostic. The Q-Prior circuit uses only standard single-qubit rotations and nearest-neighbour entangling gates, so the same protocol can be deployed on trapped-ion, neutral-atom or any other gate-based quantum platform. What matters in practice is measurement precision and two-qubit fidelity, not the specific hardware architecture.

Which combinations of existing simulation or computational methods are you considering next? What are your plans going forward?
Coveney: We are extending QIML in two directions. The first is broader applications: we are actively exploring problems in biology, where heavy-tailed cell-level variability and intermittent behaviour are pervasive, and in climate-relevant phenomena, where long-term statistical fidelity over extended rollouts is the central modelling challenge. The second is tighter QPU + GPU integration: we are developing more closely coupled hybrid workflows that bring quantum and classical resources into a single training and inference loop, in continued collaboration with LRZ and on NVIDIA's CUDA-Q framework. Interview: vs | LRZ

The image above shows a digital model illustrating how signals are transmitted or genes are activated in cells through the binding of substances. GPCRs are also involved in such processes. AI-generated graphic: Huzaifa | Adobe