An NMR instrument learns stronger signals from its own experiments

Aarhus University’s 27 October explanation spotlights a spectrometer that learns microwave pulse sequences from its own measurements. The study appeared in Science Advances in September; the experimental details below come from its March preprint.[3][2][1]

Dynamic nuclear polarisation transfers some of the stronger electron-spin alignment to atomic nuclei, strengthening the nuclear magnetic resonance signal. The timing matters: a 2019 TOP-DNP experiment at 0.35 tesla coordinated microwave pulse lengths and the gaps between them with nuclear precession. It enabled efficient transfer with much less microwave power than conventional spin locking.[4]

That earlier study also illustrates why predictions need physical checks. A simplified model with one electron and two protons initially predicted excessive signal enhancement. The authors added relaxation—the loss of spin order—and adjusted parameters to reproduce experimental behaviour. Real samples contain many surrounding nuclei, so a useful pulse sequence must work amid interactions that a small model omits.[4]

In the Aarhus preprint, a Bayesian optimiser uses past pulse settings and measured signals to predict useful experiments. Its probabilistic model balances expected improvements against poorly explored settings. Direct feedback can capture spin interactions and instrument imperfections absent from a simplified simulation.[1]

For a 2mM 4-Oxo-TEMPO solution at 80K, about −193°C, the preprint reports a 13-fold proton-signal enhancement for the best Bayesian sequence, versus 9-fold for the best of 2,000 random sequences. Both are relative to the thermal-equilibrium signal. A separate spin-lock comparison yielded about 70% higher sensitivity—the gain highlighted in the university report. The demonstration concerns frozen radical solutions; other sample types still need testing.[1][3]

Report an error or suggest a correction