Kelvin Four-Wire Sensing: Measuring Milliohms When the Wires Fight Back

The problem: your wires are bigger than your signal

Resistance of a conductor follows R = ρL/A, where ρ is resistivity (copper ≈ 1.68×10⁻⁸ Ω·m at 20 °C), L is length, and A is cross-sectional area. A 1 m round trip of 24 AWG copper (A ≈ 0.205 mm²) has R ≈ 0.164 Ω. Add two crimp or connector contacts — each typically 1–50 mΩ, and worse as they oxidize — and the parasitic path in series with your device under test (DUT) can easily reach 0.2–0.5 Ω.

A simple two-wire ohmmeter forces a known current I through the DUT and measures the voltage across its own terminals. But that voltage is V = I·(R_x + 2R_lead + 2R_contact). The meter has no way to separate the DUT resistance R_x from the wire and contact terms — it reports the whole sum. When R_x is a 100 Ω resistor, a 0.3 Ω error is 0.3% and often tolerable. When R_x is a 10 mΩ shunt or a 50 mΩ battery tab, the same 0.3 Ω is a 3,000% error. The measurement is dominated entirely by the plumbing.

• Temperature makes it worse: copper's temperature coefficient is ≈ +0.393%/°C, so lead resistance drifts ~4% over a 10 °C swing — an unstable, uncalibratable error.
• Contacts drift: fretting, oxidation, and clamp force turn contact resistance into a noisy, time-varying term that no offset subtraction can fix.

How Kelvin sensing works: separate the force from the sense

The insight, credited to William Thomson (Lord Kelvin) and his 1861 double bridge, is to use four conductors instead of two. Two of them — the force (or drive/excitation) leads — carry the measurement current I into and out of the DUT. The other two — the sense leads — tap the voltage right at the DUT terminals and route it to a voltmeter with an enormous input impedance (Z_in ≥ 10 GΩ, i.e. ≥ 10¹⁰ Ω).

Because the voltmeter draws almost no current (a few picoamps to a nanoamp), Ohm's law says the drop across the sense-lead resistance is I_sense·R_sense-lead ≈ 1 nA × 0.2 Ω = 0.2 nV — utterly negligible. The sense leads therefore report the true potential difference at the taps, and:

R_x = V_sense / I_force

The force-lead resistance still exists, and the current source must overcome it (compliance voltage), but it appears outside the two sense taps, so it never enters the voltage the meter reads. Contact resistance at the force contacts is likewise pushed outside the sensed region. The critical geometric rule: the sense contact must sit inside (or between) the force contacts, so the sensed volume of the DUT carries the full current with no parasitic series terms.

• Force leads — low impedance matters (drive current), resistance is cancelled anyway.
• Sense leads — resistance is irrelevant because I ≈ 0; what matters is that they connect physically inside the current path.

Why the current source (not just the meter) does the work

Four-wire sensing needs a known, stable current, so the front end is almost always a precision current source feeding the force leads, with the voltmeter (typically a high-resolution ADC — a sigma-delta or integrating type in a DMM) reading the sense pair. Fixing I lets you compute R directly and lets you pick I to trade signal against self-heating.

Signal is I·R_x, so a 10 mΩ DUT at 1 mA gives only 10 µV — near thermocouple-EMF and amplifier-offset noise. Raise I to 1 A and you get 10 mV, comfortably above noise, but you dissipate P = I²R = 10 mW in the DUT and far more in the leads. For an RTD the self-heating limit is the real constraint: a Pt100's error is I²R⁄S_dis, where S_dis is the self-heating (dissipation) coefficient (~2.5 mW/°C in still air). At 1 mA a Pt100 (≈ 100 Ω) dissipates 0.1 mW → ~0.04 °C error; at 10 mA it dissipates 10 mW → ~4 °C error — unacceptable. That is why RTD front-ends run ~0.1–1 mA.

• Thermal EMFs: junctions of dissimilar metals generate ~µV-level Seebeck voltages that masquerade as R. The fix is current reversal: measure at +I and −I, then average, so any DC offset (thermal EMF, amplifier offset) cancels. This is standard on µΩ meters and battery testers.
• Compliance: the source must supply V = I·(R_x + 2R_force + R_contact). Long or thin force leads can exceed the source's compliance voltage, causing the current to fall short and the reading to error out.

Sizing the measurement: currents, voltages, and error budget

A clean design procedure for a Kelvin measurement:

• 1. Set the target signal. Pick a sense voltage well above the front-end noise + offset floor. For a good instrumentation amp/ADC, keep V_sense ≥ 1–10 mV so that a 1 µV noise floor is < 0.1% of signal.
• 2. Solve for current. I = V_sense / R_x. For R_x = 5 mΩ and V_sense = 10 mV → I = 2 A. For a 100 Ω RTD and V_sense = 100 mV → I = 1 mA.
• 3. Check self-heating. Compute P = I²R_x in the DUT and ΔT = P⁄S_dis. Back off I if ΔT exceeds your accuracy (e.g. < 0.01 °C for a reference RTD).
• 4. Check force-lead compliance. Ensure the source can supply I·(R_x + R_force,total). At 2 A through 0.4 Ω of leads you need ≥ 0.8 V of headroom just for the wires.
• 5. Suppress thermal EMFs. Use current reversal (average of ±I), keep sense-lead junctions isothermal, prefer copper-to-copper.

Residual error terms to budget: sense-lead leakage (I_bias × Z_lead), input-offset voltage after reversal, current-source accuracy and stability (this maps directly onto R error since R = V/I), and ADC gain/linearity. A 6½-digit DMM on its 100 Ω four-wire range reaches ~10 µΩ resolution and ~0.01% accuracy; dedicated micro-ohmmeters push to 1 µΩ using 10 A pulses and current reversal.

Real hardware: from Kelvin clips to four-terminal shunts

Kelvin sensing shows up wherever small resistances must be trusted:

• Kelvin clips and probes: spring clips with two electrically isolated jaws — inner jaw senses, outer jaw forces — so a single mechanical bite makes the four-wire connection. Standard on LCR meters and battery-IR testers.
• Four-terminal current shunts: a precision manganin or Zeranin bar (α ≈ ±10 ppm/°C) with two heavy current terminals at the ends and two voltage-sense terminals inboard. A 50 mV / 100 A shunt is a 0.5 mΩ four-terminal resistor; because sense taps are inboard of the current lugs, torque and corrosion on the big lugs don't affect calibration. Metrology shunts hold ~0.01% over years.
• RTDs (Pt100/Pt1000) per IEC 60751: four-wire is the reference configuration; three-wire is a compromise that cancels lead R only if the two leads match, while four-wire cancels it unconditionally.
• Battery internal resistance & interconnects: EV pack testers Kelvin-probe each weld/busbar (0.05–5 mΩ) to catch bad welds; cell IR is measured four-wire, often with AC (1 kHz EIS) to separate ohmic from polarization impedance.
• On-PCB current sensing: low-value SMD sense resistors (1–10 mΩ, e.g. 2512/4-terminal packages) use a Kelvin (2-pad sense) footprint so trace and solder-joint resistance stay outside the sensed node — essential in switch-mode-supply and motor-drive current loops.

Failure modes, limits, and best practice

Four-wire sensing is robust but not foolproof. The classic mistakes:

• Sense taps outside the force taps. If a sense lead connects beyond a force contact, part of the force-lead/contact resistance falls inside the sensed region and reappears in the reading. On a four-terminal shunt this means always sensing inboard of the current lugs; on a PCB it means tapping the resistor pad, not the fat current trace.
• Swapped force/sense pairs. Wiring the sense pair as force (or vice-versa) can still "work" numerically but reintroduces lead error or overloads the high-Z input; label and route pairs carefully.
• Thermal EMF and 1/f noise dominate below ~1 mΩ. Use current reversal and, for the last decade, a lock-in / synchronous approach: drive a small AC current and demodulate at that frequency to reject DC drift entirely.
• Insufficient force current or compliance → weak signal buried in noise; check I·R_lead against source headroom.
• Cable capacitance and AC: at kHz excitation, lead inductance/capacitance and thermocouple junctions add reactive error; use guarded/twisted sense pairs and 4-terminal Kelvin fixtures rated for the frequency.
• Common-mode & guarding: for sub-µΩ work, guard the sense leads and keep the whole loop isothermal; even 1 °C across a copper-solder junction is a ~µV-class EMF that swamps a µΩ signal.

Best practice, in one line: force the current on the outside, sense the voltage on the inside, reverse the current to kill offsets, and pick I as large as self-heating allows. Do that and a $2 sense resistor or a hand-held clip measures resistances a two-wire meter can't even see.

Two-wire vs. four-wire (Kelvin) resistance measurement, measuring a 10 mΩ device through 0.2 Ω leads at 1 A

Attribute	Two-wire ohmmeter	Four-wire (Kelvin)
What is measured	R_x + 2·R_lead + 2·R_contact	R_x only (between sense taps)
Reading for 10 mΩ DUT	≈ 410 mΩ (41× error)	≈ 10 mΩ (< 0.1% error)
Lead-resistance error	Full (100s of mΩ)	Cancelled to nA × Z_lead
Wires / connections	2	4 (force pair + sense pair)
Best floor	~0.1–1 Ω useful	µΩ (with current reversal)