NMR spectrometer temperature drift

My final-year MChem project at the University of Manchester, supervised by Prof. Gareth Morris and working with Paul Bowyer, programmed a Varian NMR spectrometer to investigate a subtle source of error: the building's air-conditioning cycled on and off every few minutes, nudging the probe temperature up and down, which in turn shifted the phase and frequency of the NMR signal. Because the aircon cycle is periodic rather than random, that drift doesn't average out like ordinary noise — in a 2D spectrum it shows up as a coherent artefact, a faint diagonal running parallel to a real peak. The fix was simple once identified: randomise the order in which the experiment's increments are acquired, so the drift no longer lines up with anything. The diagram below runs the same 24-increment experiment twice — once acquired in order, once shuffled — against the same underlying temperature drift.

(Real published paper from this work: Bowyer, P.J., Swanson, A.G., Morris, G.A. “Analyzing and Correcting Spectrometer Temperature Sensitivity.” Journal of Magnetic Resonance, 152(2), 234–246 (2001). It's paywalled, so this section is built from the abstract and general NMR knowledge rather than the full methods. Prof. Morris had a memorably sharp eye for detail — he once stopped a tutorial to correct a slide that used “data” as a singular noun.)

Each cell is one t1 increment, coloured by how far the probe temperature had drifted at the moment it was measured (blue = cooler, red = warmer). Acquired in order, the drift traces a smooth ripple that survives into the spectrum as a ghost peak. Shuffled, the same 24 temperature readings land in a different order and read as noise.

  1. Aircon cycling — probe temperature drifting ±0.3°C every ~4 min
  2. Sequential run starts — 24 t1 increments acquired back-to-back
  3. Sequential 2D FT — drift is now a smooth function of increment number
  4. Result: a coherent artefact — a faint diagonal beside the real peak
  5. Randomised run starts — same 24 increments, acquisition order shuffled
  6. Randomised 2D FT — drift no longer correlates with increment number
  7. Result: drift spreads into the noise floor — the real peak stands clean
  8. Same physical drift, different acquisition order, different outcome