When computers give up, the spectrometer learns for itself

An NMR spectrometer can reveal molecular structure down to the atomic level, but weak signals limit what it can see. Aarhus University researchers use AI to let the spectrometer learn the best pulse sequence directly from the sample.

Imagine trying to shepherd a group of sugar-fuelled children through an amusement park. You cannot work out in advance exactly which combination of words, tone of voice, timing and body language will keep everyone moving in the right direction. The children react differently, they influence one another, and the larger the group, the harder it becomes to predict what will work.

At some point, the best strategy is simply to try something, see what happens and adjust.

With no other connection whatsoever, that is roughly the strategy the Aarhus researchers have applied to a very different problem: controlling large systems of atomic nuclei and electrons.

Spin is a quantum property that makes atomic nuclei and electrons behave somewhat like tiny magnets. It is the physical basis of nuclear magnetic resonance, NMR, and of MRI scanners.

The difficulty is that only a very small excess of nuclear spins point in one direction rather than the other, which produces a weak overall signal.

“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,” says Professor Niels Chr. Nielsen from Aarhus University.

Electrons to the rescue

Electrons can help because their magnetic moment is about 660 times larger than that of a proton. Their spins can therefore become much more strongly ordered in a magnetic field.

Using carefully designed microwave pulses, researchers can transfer some of this order from the electrons to the atomic nuclei.

The technique is known as dynamic nuclear polarisation, or DNP.

DNP is well established, but using it efficiently is far from straightforward.

A system of electrons and nuclei can contain hundreds or thousands of spins, all interacting with one another. That makes it extremely difficult to calculate the ideal sequence of microwave pulses in advance. On top of that, a computer model may not capture every imperfection in the real instrument.

So why build a perfect model of reality when reality itself is sitting in the laboratory?

The researchers connected a home-built DNP spectrometer to a machine-learning algorithm based on Bayesian optimisation. The system tests a pulse sequence directly on the sample, measures the result, learns from it and proposes the next experiment.

“What is new is that we let the spectrometer learn for itself how to perform the best experiments,” says PhD student Filip V. Jensen, first author of the study, which has been published in Science Advances.

The researchers tested the method on two different spin systems. In one, the algorithm found solutions comparable with some of the best existing approaches. In the more complex system, the pulse sequence found by the algorithm produced around 70 per cent greater signal enhancement than the reference method used by the researchers.

Each optimisation took from just under two days to almost five days. But once an effective sequence has been found, it can be reused. And greater sensitivity means fewer repetitions are needed to obtain a useful signal, reducing the total measurement time.

A step towards better quantum sensors

The study is a proof of concept using model systems. It is fundamental research, but Niels Chr. Nielsen sees a wider perspective.

The same approach could become a tool for developing better magnetic-resonance-based quantum sensors, where large and complex spin systems have to be controlled without first being simulated in every detail.

In the longer term, he points to possible applications in advanced laboratories, industrial process monitoring, environmental measurements and food analysis.

Other analytical methods may be more sensitive if the only aim is to detect whether a tiny amount of a substance is present.

NMR offers something different: detailed information about how molecules and materials are structured.

Niels Chr. Nielsen puts it this way:

“If you put my head into an optical sensor, you would probably mostly learn that I do not have much hair. But in an MRI scanner, you have a chance of seeing what is inside.”

---

---