The A-D converter, and why we had to build one.
Precision thermometry depends on converting very small voltages without adding significant error. In 2004, no available A-D converter could do what we needed, so we designed one. More than twenty years later, the same approach is still in use.
The 2004 problem
The instruments then working at or below 1 mK fell into two camps. Some were chopped-DC potentiometric designs; the others were AC resistance bridges. We wanted the microK to handle thermocouples as well as platinum resistance thermometers and thermistors, and that ruled the AC bridge out, because an AC bridge has no sensible way to measure a DC voltage from a thermocouple.
That left DC, and DC instruments of the day used integrating converters. Their linearity is limited to a few parts per million by dielectric absorption, the reluctance of an integrating capacitor to give back quite all of the charge it was given. The best commercially available part managed 3 ppm. Good, and not good enough for the specification the project had set.
The other option was to accept the converter and correct for its behaviour in software. We were not willing to do that. A correction is a model of the part you happen to have measured, and it stops being true as the part ages, warms up or is replaced by the next one off the reel.
What we built
A multi-bit adaptation of the sigma-delta technique, developed with NPL and Metron Designs, reaching better than 0.5 ppm linearity. Every instrument in the microK family, the milliK, and the scanners that expand them, has been designed from it.
Substitution, and what it removes
A resistance bridge has to compare an unknown resistance against a reference. The conventional approach passes a common current through both and compares the two voltages it develops, which means the two signals travel different paths to the converter and pick up different errors on the way.
We switch instead. The device under test and the reference are brought in turn to the same measurement point, so the converter sees one signal path and one set of imperfections. Gain error, offset and drift then appear identically in both readings and divide out of the ratio. The measurement is drift-free by construction, which is a much weaker demand on the electronics than asking them not to drift.
No relays anywhere
The switching is solid-state throughout. Mechanical relays are the component in an instrument like this most likely to fail, and they were the wrong answer in 2004 for the same reason they are the wrong answer now.
It did cost us something. PhotoMOS devices have a finite off-resistance, and at the levels these instruments work at, leakage through a switch that is supposed to be open lands directly on the measurement. Building the microsKanner meant designing a buffered switching arrangement of our own to remove that leakage, which is why ninety extra channels cost the microK nothing in performance.
0.1 ppm, and a picosecond
Taking the platform into primary standards work in 2009 meant matching AC bridges outright, and that meant 0.1 ppm. Two things had to improve together. Converter linearity came down to timing jitter, which had to fall below a picosecond, the time an electrical signal takes to travel about a third of a millimetre. Separately the input noise had to drop far enough for the better linearity to be worth having, and no published approach came close, so that too had to be designed: a parallel analogue processing technique that remains ours alone.
The microK² ELITE and the microK² 60 were designed from that work, fifteen years on.