Water vapour lines and what they do to a transmission spectrum
Why atmospheric absorption lines throw spikes into a transmission ratio, why purging is the real fix, and how to reject the lines you could not purge away without also deleting your sample.
THz-TDS transmission calculator
Everything described here, worked out in your browser. Nothing is uploaded.
Why a line becomes a spike
Rotational transitions of water vapour cut narrow, very deep notches into any terahertz spectrum measured in air — at 1.10, 1.16, 1.21, 1.41, 1.67, 1.72, 2.04, 2.16, 2.46, 2.58, 3.09 and 3.25 THz among many others. At the bottom of one of these lines, almost nothing reaches the detector.
Now divide the sample spectrum by the reference. At a line, both numbers are nearly zero. The ratio of two nearly-zero numbers is dominated by whatever differs between them, and what differs is the humidity in the beam path at the two moments you measured. A purge that improved slightly between the reference and the sample scan throws the ratio upward; a purge that worsened throws it down. Either way you get a spike several times the height of the surrounding curve, at a frequency where your sample is doing nothing at all.
These spikes are narrow, sharp and repeatable on reprocessing, which makes them look exactly like real absorption features. They are not.
The real fix is the measurement
Purge the beam path with dry nitrogen, and give it long enough to settle before you start. Measure the reference and the sample as close together in time as you can, so the humidity has less opportunity to drift between them. If you can only purge partially, at least keep the purge steady: a constant residual humidity divides out, a changing one does not.
Rejecting what you could not purge
Where the lines are already in the data, they can be removed, but the removal has to be able to tell a narrow atmospheric line from a broad sample absorption. A hard-coded list of line frequencies is the obvious approach and the wrong one: it misses instrument artefacts and it assumes a line width that depends on your pressure and resolution.
A local test works better. Compare every frequency bin with a running median of the same spectrum, taken over a window much wider than a rotational line and much narrower than anything a sample does — around 0.15 THz is a reasonable choice. Anything sitting far below its own neighbourhood is a notch. A broad absorption band drags the median down with it and survives; a narrow dropout does not.
Two thresholds are better than one. Seed a line where the deficit exceeds the full threshold, then grow the mask outwards while the deficit is still over half of it. That covers a deep line out to its wings, where the beam is still too weak to trust, without over-trimming a shallow one. Whatever you do, report how many points you deleted. A cleaning step that does not tell you what it removed is not a measurement step.
