Terahertz Transmission

The ratio is the measurement. Both scans are transformed and the sample spectrum is divided by the reference, so everything common to the two — the emitter, the detector, the optics, the water lines — cancels. That only holds if nothing else changed between them. A realigned beam or a different purge level shows up as fake structure in the transmission.

Dynamic range is the honest limit. The noise floor is estimated from the quiet stretch before the pulse arrives. Where the reference drops within about 10 dB of that floor, the ratio is noise divided by noise and the transmission there means nothing, however smooth it looks. The usable-bandwidth figure is where that happens.

Water lines are not your sample. Rotational lines of water vapour cut narrow, very deep notches into both scans. At the bottom of one, almost nothing gets through the air, so the ratio is one nearly-zero number over another and the smallest difference in humidity or purge between the two measurements throws a spike metres high. Those spikes look like sharp features and they are not. The tool removes them by comparing every bin with a running median of the same spectrum, about 0.15 THz wide, and dropping whatever sits more than the set depth below its own neighbourhood. It is a local test, so a broad absorption band moves the median with it and survives, while any sharp dropout goes — water lines, but also a bad bin or a notch filter you forgot about. Turn it off to see the raw ratio, and raise the depth if it is eating something you want to keep. The status line says how many points went. The real cure is still to purge, or to run both scans at the same humidity.

Field or power, and the decibel that joins them. The ratio the tool computes is a field amplitude: the sample spectrum divided by the reference spectrum, both being amplitudes. Power transmission is its square, which is what a bolometer or a power meter would report. In decibels the two coincide — 10 log₁₀|t|² is 20 log₁₀|t| — so switching between field and power changes the linear and per-cent plots but leaves the dB plot exactly where it was. That is not a bug; it is the reason decibels are quoted in the first place. The exported CSV carries all three columns whichever one is on screen.

Windowing is a choice, not a default. Cutting the record before an internal echo removes the Fabry–Pérot ripple, at the cost of frequency resolution. Leaving the echo in keeps resolution and gives you the ripple. The tool does whichever you ask and tells you the resolution you ended up with.

Index and absorption assume a simple slab. With a thickness, the tool uses n = 1 + cΔφ/ωd and takes the Fresnel losses off the amplitude. That is the standard thick-sample extraction and it ignores multiple internal reflections, so it is right when the echo is outside your window and approximate when it is not. Thin or strongly absorbing samples need a proper iterative fit, not this.