When simulations fall short, this spectrometer learns for itself

Introduction

Nuclear magnetic resonance (NMR) and its medical cousin, MRI, are powerful tools for probing molecular structure, but their sensitivity is limited by the tiny net magnetisation of nuclear spins. Researchers at Aarhus University have tackled this bottleneck by letting the spectrometer itself learn the best dynamic nuclear polarization (DNP) pulse sequence. Using a Bayesian‑optimization algorithm that evaluates each experiment in real time, the instrument discovers pulse patterns that dramatically enhance signal strength, sidestepping the need for exhaustive simulations of complex spin systems.

Why traditional modelling falls short

In DNP, microwave pulses transfer the large magnetic moment of electrons (about 660 times that of a proton) to nearby nuclei, amplifying the NMR signal. Designing the optimal sequence of pulses is a daunting task: a typical sample contains hundreds to thousands of interacting spins, and imperfections in the hardware are hard to capture in a computer model. As Professor Niels Chr. Nielsen notes, "NMR has the same disadvantage as MRI scanners: it is relatively insensitive. The samples need to contain a high concentration of whatever you want to analyse." Consequently, relying solely on simulations can lead to sub‑optimal performance when the experiment is run on a real instrument.

The self‑learning spectrometer

The Aarhus team built a home‑made DNP spectrometer and coupled it to a machine‑learning loop based on Bayesian optimization. The workflow is straightforward:

1. Propose a candidate microwave pulse sequence.
2. Execute the sequence on the actual sample.
3. Measure the resulting NMR signal.
4. Feed the signal back to the optimizer, which updates its model and suggests the next sequence.

This closed‑loop approach treats the sample and the hardware as part of the optimisation problem, allowing the algorithm to adapt to real‑world quirks that a simulation would miss.

Results and performance

The researchers tested the method on two model spin systems. In the simpler case, the AI‑derived sequences matched the best existing handcrafted approaches. For the more complex system, the algorithm produced a pulse sequence that delivered roughly 70 % greater signal enhancement than the reference method used by the team. Each optimisation run required between two and five days of continuous experimentation, after which the discovered sequence could be reused for future measurements, reducing the number of repetitions needed for a high‑quality spectrum.

Implications and future directions

Although the study remains a proof‑of‑concept on model compounds, it points to a broader paradigm shift for quantum‑sensor development. By allowing instruments to learn directly from the physical system, researchers can bypass the labor‑intensive task of modelling every interaction. Potential applications include:

• Advanced quantum sensors that must control large, entangled spin ensembles.
• Industrial process monitoring, where rapid, sensitive detection of chemical changes is valuable.
• Environmental and food analysis, where detailed molecular information can complement more sensitive but less specific techniques.

NMR’s unique strength lies in its ability to reveal structural details, not just the presence of a substance. As Nielsen puts it, an MRI scanner can show what lies inside a body, whereas a simple optical sensor would only tell you about surface features. The self‑learning spectrometer enhances that capability by making the underlying measurement more robust and efficient.

Practical take‑away

For laboratories facing the trade‑off between simulation effort and experimental fidelity, integrating Bayesian‑optimization loops into existing spectrometers offers a pragmatic path forward. The approach requires modest hardware modifications—primarily a programmable pulse generator and a data acquisition interface—but can yield substantial gains in sensitivity and throughput without the need for exhaustive theoretical modelling.

--- This article summarises the work of Filip V. Jensen and colleagues, published in Science Advances (2026), and reflects on its significance for the future of NMR and quantum‑sensor technologies.