Quantum Computing · Cryogenic Thermometry
This is the checklist every quantum hardware procurement team should keep on hand — written from the supplier-quality side of the table.
A superconducting qubit is only a qubit when thermal excitations are negligible. A 5 GHz qubit corresponds to a photon energy of roughly 240 mK; to make thermal excitation irrelevant, the processor has to run an order of magnitude colder — 10 to 20 mK, exactly the base temperature of a dilution refrigerator.
At that point the thermometer itself becomes the most delicate component in the system. The cooling power of a large dry dilution refrigerator's mixing chamber is nanowatts to microwatts. Every wire, every resistor, every chip crossing a temperature stage leaks heat inward — and every measurement you make adds a little more. "How cold is the fridge?" sounds like a simple question. A traceable, correct answer demands metrology discipline that most supply chains have never been asked to master.
This article walks through the six challenges of dilution-refrigerator thermometry and closes with a supplier checklist — what to demand from thermometer and wiring vendors before you pay, from the supplier-quality side.
Modern dry dilution refrigerators (Bluefors XLD series, Oxford Instruments Triton, FormFactor systems) cool with a ³He/⁴He mixture across discrete temperature stages:
| Stage | Typical temperature | What lives there |
|---|---|---|
| 50 K / 3 K shields | 50 K, ~3–4 K | Radiation shielding, first-stage heat anchoring, amplifiers |
| Still | ~700–800 mK | The pumped ³He bath where evaporative cooling happens |
| Cold plate | ~100–150 mK | Intermediate heat anchoring, filtering |
| Mixing chamber (MXC) | Base 10–20 mK | The processor chip. Below ~0.87 K, cooling occurs at the phase boundary |
Cooling power tells the whole story: a large dry unit delivers several hundred microwatts at 100 mK — and only nanowatts at base temperature. Every pitfall below traces back to that collapse.
No single sensor spans 300 K down to 10 mK, so refrigerators stack solutions by stage:
Germanium has the best sensitivity below 1 K but is brittle, slow, and notoriously magnetically sensitive; RuO₂ thick-film chips are robust with wide range but each one needs individual calibration; Cernox has the best magnetic-field tolerance of the bunch. Selection isn't the hard part — getting every single unit's calibration solid is.
Resistive thermometry works by driving current and measuring voltage. Driving current generates heat — and at base temperature, power is the scarcest resource in the system.
Run the numbers: 1 mV of excitation across a 2 kΩ RuO₂ chip dissipates half a nanowatt. The total cooling power of the mixing chamber at 15 mK may be on the same order. At that point the thermometer is no longer measuring the bath — it is heating the bath, and itself. The industry rule of thumb: sensor dissipation should stay under 1% of the stage's cooling power. Excitation voltages get pushed down to tens of microvolts, AC resistance bridges (Picowatt AVS-47B, Lake Shore Model 372), four-wire connections, and aggressive filtering are all non-negotiable.
This error is silent: the thermometer reads a few percent warm, the curve looks stable — and the "base temperature" in your paper or datasheet is partly an artifact of your own measurement power.
The mixing chamber plate is not isothermal. With loom and sample package attached, a few millikelvin of variation across the plate is normal. So "the fridge is at 12 mK" really means: the point where the thermometer sits is at 12 mK.
There is a subtler layer. At millikelvin temperatures, electron–phonon coupling in metals weakens dramatically. The processor is always being driven, and under drive the device's electron temperature rises above the lattice temperature at its mounting position. The thermometer reports the equilibrium of its own electrons with its own lattice; the qubits on the chip obey a different one. Both temperatures are real — the dashboard shows only one number.
Operational discipline: instrument every stage and every critical location; do in-situ calibration where possible; and be honest about what the number actually refers to. A buyer who takes "mixing chamber: 12 mK" at face value has bought a number, not a measurement.
Most quantum-computing refrigerators are not just cold — they are cold inside a magnet, because superconducting and spin qubits are tuned with static fields. Resistive thermometers drift in a field. Germanium is famously sensitive; RuO₂ shifts noticeably above 1 T; Cernox is among the most field-tolerant, which is why it survives in magnet systems.
So a zero-field calibration curve is only partially valid inside a solenoid. Buyers should demand magnetoresistance data over the target field range — or arrange their own in-situ comparison. Not specifying the end-use magnetic field at order time is a more upstream specification error than most people realize.
Every wire entering the cold stages conducts heat from room temperature. The standard countermeasure is a materials ladder: phosphor-bronze or cupronickel (CuNi) looms on the warmer stages; below roughly 9 K, superconducting NbTi — a superconductor carries signal while conducting almost no heat, which makes NbTi the default for millikelvin and qubit signal wiring (including Thermocoax filtering cable with embedded NbTi).
But a loom is not a catalog item. It is a component with real quality attributes: continuity and insulation, heat-anchor points at every stage, dressing quality, filter cutoff consistency, and the length tolerances that decide where the anchors land. One poorly anchored loom warms the fridge and dirties every signal — and because thermometry wiring shares space with qubit control lines, one bad loom can corrupt both your temperature data and your experimental data.
The question that separates "buying a component" from "buying a product" is: what is the traceability chain of this thermometer's calibration?
ITS-90 only extends down to 0.65 K. Below that, the reference is the Provisional Low Temperature Scale of 2000 (PLTS-2000, 0.9 mK to 1 K), built on the melting pressure of ³He. So when a vendor hands you a calibration certificate for a 10 mK thermometer, the questions worth asking are:
Finally, the dynamic environment: cooldown from 300 K to 10 mK takes a day or two, with condensation risk, vibration, and electromagnetic interference from qubit control pulses along the way. Measure while experiments run and you read a sliding average of a noisy world. Even "base temperature" implicitly assumes the fridge is at rest — the moment the processor is driven, the device's effective temperature exceeds anything any thermometer reads.
Nobody can eliminate these effects. The difference between a credible supplier and a filler supplier is whether they have characterized them, documented them, and stated the boundaries.
The ecosystem is narrow and specialized: calibration sensor vendors (Lake Shore, Scientific Instruments), primary-thermometry vendors (Aivon's CBT line), and integrators selling their own thermometry and wiring assemblies (Bluefors, Oxford Instruments, FormFactor), plus a handful of custom loom and filter shops. Many are genuinely good. But most small shops will not survive one serious audit — which doesn't necessarily mean their product is bad, only that their quality system isn't yet a system. The buyer has to bring it, because the supplier probably doesn't have one.
That is exactly where supplier verification earns its price.
| # | Check | What to demand from the vendor |
|---|---|---|
| 1 | Traceability | Which scale (PLTS-2000 / ITS-90)? Which laboratory? ISO/IEC 17025 accreditation or written evidence? |
| 2 | Curve completeness | Full R–T curve with uncertainty budget, or isolated points? |
| 3 | Thermal-cycling stability | Drift data over at least three 300 K↔operating cycles. |
| 4 | Self-heating | Measured dissipation at recommended excitation. Numbers. |
| 5 | Magnetoresistance | Characterization over the final operating field range — or an honest "we don't have it." |
| 6 | Response time | Thermal time constant at operating temperature, including the recommended wiring and filtering. |
| 7 | Package & interface | Chip carrier, lead materials (NbTi vs CuNi), pinout, compatibility with standard anchoring hardware. |
| 8 | Cross-check records | Has the vendor cross-verified against a reference unit? Can you see the data? |
| 9 | Re-calibration | Turnaround, cost, drift criteria. |
| 10 | Batch consistency | Is a multi-unit order from the same calibration family? Is unit-to-unit spread documented? |
Acceptance testing is cheap insurance: compare the new sensor against a trusted reference in a 77 K liquid-nitrogen bath; add a 4.2 K liquid-helium point if you can; run a thermal cycle; do one excitation sweep to map self-heating — and only then let it report your base temperature. Two hours of bench work catches most certificate problems.
An article that cannot be proven wrong cannot be proven right. So it closes with three testable claims — the evidence isn't in yet, and the claims are on the record. When the time comes, come back and settle accounts:
If I'm wrong, the correction goes into the update log below.
The ten questions above are free. Getting answers out of a real factory takes ten years of sensor-industry experience and an auditor's license — I hold both. If you are about to wire a prepayment to a temperature-sensor or wiring vendor, see Supplier Verification Services: remote desktop verification, live video audits, and a bilingual scorecard report with a plain GO / CAUTION / NO-GO verdict.
About the author: 15 years in the temperature-sensor industry; CCAA-certified ISO 9001 quality-management-system auditor.
Technical values (stage temperatures, cooling powers, scale definitions, product categories) reflect standard cryogenic engineering practice. Readers deploying critical systems should verify specific parameters against their own hardware documentation and calibration certificates.