Selecting a reference thermometer

Everything you calibrate inherits the reference. Its certificate uncertainty, its drift since that certificate, and the readout you measure it with all propagate into every result you issue, and no amount of care downstream recovers what was lost there. So the reference is chosen against the uncertainty you have promised to deliver, not against what looks impressive in a catalogue.

How good does it need to be

The usual guidance is that the reference should contribute a third or a quarter of the target uncertainty, which keeps its variance contribution to roughly 11 % or 6 % and leaves room for the source, the readout and repeatability.

If the arithmetic says you need a reference better than 0.03 °C, stop and check the rest of the chain first. A 0.025 °C reference in a source with 0.2 °C of gradient is money spent on the wrong term.

The families, and where each stops

• Standard platinum resistance thermometer (SPRT): strain-free high-purity platinum meeting the ITS-90 criteria. Millikelvin uncertainty, and mechanically fragile enough that a single knock can end its calibration. For fixed-point and top-tier comparison work.
• Reference-grade industrial PRT: a rugged sheathed sensor selected for low drift, typically calibrated to 0.01–0.03 °C. This is the right answer for most calibration laboratories and almost all validation work.
• Thermistor: very high sensitivity over a narrow near-ambient span, excellent resolution, poor interchangeability and a strongly non-linear characteristic. Good for a fixed narrow range, awkward as a general reference.
• Thermocouple: the only practical option above about 660 °C, but inhomogeneity makes it a poor reference wherever a PRT will reach. Below 400 °C, choosing a thermocouple as your reference is choosing a worse instrument.

The readout is part of the reference

A reference PRT is a resistor. What you actually own is the sensor and the instrument that measures it, and the instrument's specification enters the budget on equal terms. A Pt100 changes 0.3908 Ω per °C at 0 °C, which means 1 mΩ of readout resolution or accuracy is 2.56 mK. A bridge resolving 1 mΩ is comfortably matched to a 0.01 °C sensor; a handheld meter resolving 10 mΩ gives you 25.6 mK of quantisation alone and has quietly halved the value of the sensor you paid for.

Sense current belongs in this decision too. The measuring current dissipates power in the element, and the resulting self-heating depends on the medium — under a millikelvin in a stirred bath, tens of millikelvin in still air. Calibrate and use at the same current, in a comparable medium, and the effect largely cancels; do otherwise and it becomes an uncorrected bias.

Stability beats the certificate number

A reference with a 0.01 °C certificate that drifts 0.05 °C a year is worse than one with a 0.03 °C certificate that holds. The certificate tells you where the thermometer was on one day; drift tells you where it is now, and only history reveals it.

• Check the ice point regularly — monthly is not excessive for a working reference — and record it. A shift at the ice point is usually the first sign of strain or contamination.
• Keep an as-found history across calibrations and set the interval from it rather than from a default year.
• Treat any mechanical shock as an event: check the ice point before trusting the thermometer again.
• For the lowest uncertainties, a triple point of water cell turns the ice-point check into a genuine fixed-point measurement.

A tolerance class is not a calibration. IEC 60751 class A is ±0.15 °C at 0 °C and ±0.35 °C at 100 °C — a manufacturing limit on an uncalibrated sensor, not a measurement uncertainty. Class AA is tighter at ±0.10 °C at 0 °C and is still not a certificate.

Matching the reference to the range

A reference calibrated from 0 to 100 °C is evidence about 0 to 100 °C. Using it at 150 °C means extrapolating its characteristic, which is exactly the assumption calibration exists to remove. Have it calibrated across the range you work in, with enough points to support the interpolation you intend, and say on your own certificates which range the reference covered.