Thermal EMF, the unwanted Seebeck voltage that appears at every junction of dissimilar metals, can be eliminated by sourcing +I and -I through the same circuit and averaging: V = (V_forward − V_reverse)/2, because the thermal offset is polarity-independent while the wanted IR drop flips sign [S1][S3].
For sub-milliohm shunts and µV-level calibration work, a single forward measurement can be off by tens of microvolts; current-reversal averaging typically knocks that error to <0.1 µV in a bench setup [S1][S3]. The technique is now standard in 6.5-digit and 8.5-digit DMMs and in any precision current-sense path where µV drift matters [S1][S3].
Why the trick works: the math of the Seebeck offset
A thermal EMF E_th at a junction is governed by the relative Seebeck coefficient of the two conductors, the temperature difference ΔT, and the integral form: E_th = ∫(Q_b − Q_a) dT, which for small ΔT collapses to the familiar Q_ab · ΔT [S5][S6]. It is critical that Q_ab is a property of the two materials, with copper-to-copper under 0.2 µV/°C, copper-to-gold about 0.3 µV/°C, copper-to-Kovar near 40 µV/°C, and copper-to-copper oxide over 1000 µV/°C [S3].
When you force current +I, the meter reads V+ = I·R + E_th; when you reverse to -I, it reads V- = -I·R + E_th. The thermal term does not flip sign because it is driven by the temperature difference at the junction, not by current direction [S1][S3]. Subtracting and dividing by two gives (I·R), the value you actually want [S3]. This is the algebraic heart of the so-called "offset-compensated ohms" or "delta" mode on bench DMMs [S1].
What this looks like on a real current-sense resistor
In a low-side current-sense path, the shunt's element-to-termination joints are the dominant thermal EMF generators, and even a 1 °C gradient there can inject microvolts into a µV-class signal [S4][S8]. Vendor demonstrations on foil current-sense resistors like the CSM2512F and CSM3637F use current-reversal averaging to show residual thermal EMF in the sub-µV range after the technique is applied, against a plain metal-strip shunt that can swing several µV/°C [S4].
For industrial 4-20 mA loops, a 250 Ω sense resistor sitting on a warm PCB next to a power stage can easily develop 0.5-2 µV of thermal EMF; that is 2-8 ppm of a 250 mV full-scale drop, which is exactly the region where current-reversal averaging is worth the extra settling time [S4][S7]. Differential measurement across the shunt is the practical implementation, with the meter measuring V+ and V- on the same pair of input terminals [S7].
Selection criteria: when reversal averaging is worth using
Use the technique when three things are true: the resistor value is in the milliohm-to-ohm range, the measurement bandwidth is DC or near-DC, and the thermal gradient at the junction changes slowly enough that E_th is stable between the forward and reverse samples [S1][S5]. Skip it when you need >1 kHz of true bandwidth, when the current source cannot reverse cleanly, or when the thermal EMF is changing faster than your reversal cycle (e.g. during inrush) [S5][S7].
Material choice still matters even with the technique in play. Best practice pairs the shunt and lead materials to keep Q_ab small, and connects to copper binding posts (rather than nickel-plated brass) so that the residual E_th, after averaging, is negligible rather than merely halved [S3][S5]. Inside a precision DMM, the same logic is why low-thermal binding posts and matched copper-to-copper internal wiring are used in metrology labs [S3].
Main options compared: reversal averaging, AC excitation, and chopper stabilization
Three practical techniques compete for the same µV-level error budget, and the right choice depends on bandwidth, equipment, and ambient stability. Current-reversal averaging is the simplest and works on any 4-wire ohms-capable DMM [S1]. AC excitation, used in many modern handheld ohmmeters, modulates the test signal above the thermal EMF bandwidth and rectifies synchronously; it is faster but needs an AC-capable source [S3]. Chopper-stabilized amplifiers and Kelvin-connected switches attack the same problem at the front end and are standard in high-end data-acquisition boards [S6][S7].
On bandwidth, reversal averaging tops out near the reciprocal of the cycle time (typically a few hundred ms for a 1 NPLC DMM); AC excitation reaches hundreds of Hz; chopper-stabilized front ends sit in between and are common in thermal-imaging reference instrumentation and similar front-end signal chains. On equipment cost, reversal averaging needs only a reversible current source; the other two are built into the instrument. On residual error after correction, all three can hit sub-µV in a controlled bench, but reversal averaging degrades fastest when the thermal gradient is changing during the cycle [S1][S3][S7].
Failure modes and what to watch in the field
The biggest practical failure is that E_th drifts between the +I and -I samples, so the two measurements do not see the same thermal offset and the average is incomplete; in practice you must let the junction temperature settle for seconds to minutes before the first sample, and the cycle time must be short compared to the drift rate [S2][S3]. The second failure is leakage: at very low I·R, the meter input bias current itself generates a voltage across the source impedance, and it is not cancelled by reversal averaging [S6].
Connector plating is a subtle one. Gold-plated copper binding posts have Q_ab(Cu-Au) ≈ 0.3 µV/°C, which seems small, but the polarity of each junction matters: a chain of five Cu-Au-Cu-Au-Cu junctions has some cancellations and some additions depending on the sign convention, so the naive "more contacts = more error" rule is wrong [S2][S3]. Real metrology labs use low-thermal copper binding posts and keep the loop symmetric, with the + and - paths as thermally identical as possible [S2][S3].
Standards, sourcing, and how the technique is documented
Reverse-current averaging is the technique described in Keysight's published guidance for thermal EMF error cancellation, which states explicitly that the system "supports sourcing both positive and negative current, enabling forward and reverse measurements required for error cancellation" [S1][S9]. The same approach is covered in Cal Lab Magazine's Metrology 101 article (Martin L. Kidd, Fluke), which gives the cancellation formula V = (+ Forward − Reverse)/2 and pairs it with the Seebeck coefficient table referenced above [S3].
NI's switching documentation and Vishay Precision Group's current-sense videos align with the same practice, treating thermal EMF as a material-pair phenomenon best handled by matched junctions plus reversal averaging, not by some exotic compensation algorithm [S6][S8]. For a closer look at how other low-voltage measurement chains handle junction errors, the same Seebeck logic drives thermocouple reference-junction compensation, and the same reversal-style tricks are used inside calibration and test apparatus and precision lamps and light fittings used as radiometric references. The related metrology of non-contact temperature measurement follows the same offset-cancellation logic on its reference detector.
Track two signals over the next quarter: (1) more 8.5-digit DMMs exposing delta-mode current reversal as a first-class SCPI command rather than a hidden routine, and (2) more integrated analog front ends with on-chip ±I excitation and built-in averaging, following the Keysight-style architecture [S1]. Either shift will move current-reversal averaging from a metrologist's trick into a default feature on mid-range bench instruments.