Measuring temperature with thermocouples
Why a thermocouple never measures one temperature, what cold-junction compensation is actually doing, how to choose a type, and where the error really comes from.
A thermocouple does not measure temperature
It measures a difference, and almost everything that goes wrong with thermocouples follows from forgetting that one sentence.
Join two dissimilar metals and a voltage appears that depends on the temperature gradient along the wires — the Seebeck effect. To use it you need two junctions: the one you push into the oven, and the one where the thermocouple wire meets the copper of your instrument. What a meter reads is the difference between them. There is no arrangement of wires that gives you an absolute reading, because the effect itself is differential.
So a thermocouple is really a temperature difference sensor that you turn into a thermometer by knowing what the second junction is doing. That second point is the cold junction, and dealing with it is the whole practical problem.
Why the tables assume an ice bath
Every published thermocouple table — the NIST ITS-90 reference functions this site uses included — gives voltage against temperature for a reference junction held at exactly 0 °C. Historically that was literal: a flask of crushed ice and water, which sits at 0 °C reliably enough to calibrate against.
Your reference junction is a screw terminal on a circuit board at room temperature. It is not at 0 °C, so it makes its own contribution, and that contribution has to be accounted for before the table means anything.
Vmeasured = E(thot) − E(tcold)
Cold-junction compensation is the fix, and it is less mysterious than the name suggests. Measure the terminal block's temperature with an ordinary sensor — a thermistor or a small semiconductor part, which works fine because that point is near ambient — look up the EMF it would have produced, add it back, and then read the total off the table as though the reference really had been in ice.
Every thermocouple instrument does this internally. It is why a thermocouple amplifier chip has a temperature sensor on the same die, and why the datasheet tells you to keep both terminals at the same temperature and close to that chip.
What it costs to skip it
Worth doing concretely, because the error has a signature you can learn to recognise. Take a type K probe in a 300 °C oven with the instrument on a bench at 25 °C:
| Table EMF at 300 °C | 12.209 mV | What the tables say, against an ice bath |
|---|---|---|
| Table EMF at 25 °C | 1.0002 mV | What the cold junction contributes |
| What the meter reads | 11.208 mV | The difference between them |
| Read without compensating | 275.78 °C | 24.2 °C low |
Computed at build time from the ITS-90 reference function for type K.
The error is close to the reference temperature itself — not exactly, because the curve is not perfectly straight, but close enough to be diagnostic. A thermocouple reading that is low by about room temperature, and that drifts as the room warms through the day, is an uncompensated thermocouple. It is the first thing to check when a reading looks plausible but wrong.
The thermocouple calculator shows both figures side by side so you can see the gap for your own type and reference temperature.
Choosing a type
There are more types than anyone needs. These five cover almost all work outside furnaces and calibration labs.
| Type | Alloys | Range | Output at 100 °C | Sensitivity | Choose it for |
|---|---|---|---|---|---|
| Type K | Chromel (+) / Alumel (−) | -270 to 1 372 °C | 4 096 µV | 40.5 µV/°C | The default general-purpose choice: wide range, cheap, everywhere |
| Type J | Iron (+) / Constantan (−) | -210 to 1 200 °C | 5 269 µV | 51.8 µV/°C | Higher output than K; common on older plant and plastics machinery |
| Type T | Copper (+) / Constantan (−) | -270 to 400 °C | 4 279 µV | 40.7 µV/°C | Cryogenic and sub-zero work; excellent in moist environments |
| Type E | Chromel (+) / Constantan (−) | -270 to 1 000 °C | 6 319 µV | 61 µV/°C | The highest output of the common types — best resolution per bit |
| Type N | Nicrosil (+) / Nisil (−) | -270 to 1 300 °C | 2 774 µV | 26.8 µV/°C | A drift-resistant replacement for K where stability matters |
Output and sensitivity computed from the reference functions, not copied from a datasheet.
In practice: type K unless you have a reason otherwise. It is the default on nearly every instrument, the cheapest, and the easiest to replace at short notice — which matters more than it sounds when a probe fails in the field.
Reach for type T below zero, where it stays well behaved and resists moisture. Reach for type E when you are short of resolution, since it produces the most microvolts per degree of any of these. Reach for type N when the instrument will sit at high temperature for years and drift matters more than convenience. Type J is mostly inherited rather than chosen — if the plant already runs it, match it.
Why the signal is so awkward
Type K gives about 40.5 µV per degree. That is a small number, and it drives most of the front-end design.
Feed a thermocouple straight into a bare 12-bit ADC on a 3.3 V reference and one count is worth about 806 µV — roughly20 °C per count. The whole useful range of the sensor lands inside a handful of counts. This is why thermocouples are never connected directly to a general-purpose ADC, and why dedicated amplifier and cold-junction chips exist.
It also means that microvolt-scale imperfections matter. Amplifier offset drift, thermoelectric voltages at solder joints in the input path, and common-mode noise picked up on a long cable all land in the same range as the signal. Keeping the input pair tight together, guarding the terminal block from draughts, and filtering aggressively are not fussiness — a few microvolts is a degree. Thesensor front-end calculatorcovers sizing that stage.
The wiring mistakes
More thermocouple problems are wiring than physics.
Do not extend with copper. Every junction of dissimilar metals in the circuit produces its own EMF. Splice copper onto type K partway along and you have created a second thermocouple at the splice, whose contribution depends on the temperature there. Use matching thermocouple wire, or the cheaper alloy-matched extension gradecable sold for exactly this.
Watch the polarity. Thermocouple connectors are polarised and the colour codes differ between standards, which makes reversal easy. The symptom is unmistakable once you know it: the reading moves thewrong way as the process heats up.
Keep both terminals at one temperature. Compensation assumes a single cold-junction temperature measured by one sensor. Put the terminal block in a gradient — beside a hot regulator, in the path of a fan, half in sunlight — and that assumption fails silently. The reading is wrong by however much the two terminals differ.
Mind what the probe is touching. A thermocouple reports its own junction temperature, not the thing you care about. A probe resting against a cool wall, or hanging in air above a surface, reads the compromise between them. Sheathed probes also have thermal mass, so a fast transient can be over before the sensor has caught up.
Accuracy, honestly
The conversion is exact; the sensor is not. Thermocouple wire carries a tolerance class — commonly ±1.5 °C or ±0.4% of reading for class 1 type K, whichever is larger — and that is the wire as manufactured, before anything happens to it. Ageing, oxidation, contamination and mechanical work all shift a thermocouple away from its table, and the drift is not recoverable by recalculating; it needs recalibration or replacement.
Which is the honest case for using something else when you can. Aplatinum RTD is markedly more accurate and more stable over the range where both work, and anNTC thermistor is cheaper and more sensitive still over a narrow span near ambient — which is precisely why a thermistor is usually the part measuring your cold junction.
Thermocouples earn their place on range, ruggedness, size and speed: they reach temperatures that would destroy an RTD, tolerate vibration and thermal shock, can be made fine enough to respond in milliseconds, and cost very little. Choose one for what it is good at, and do not expect precision it was never going to give you.