S-Parameter De-Embedding and Fixture Removal

When measuring a high-speed component on a Vector Network Analyzer (VNA) — whether it is a backplane connector, an IC package substrate, an AC coupling capacitor mounting pad, or a test socket — you cannot connect precision coaxial cables directly to the microscopic internal pins. You must route the test signals through a physical test fixture: coaxial SMA or RF probe launches, board breakout vias, and PCB lead-in traces.

The VNA measures everything between its reference planes as a single monolithic network: the instrument cables, the PCB launch transitions, the lead-in traces, the device under test (DUT), the lead-out traces, and the return connectors. At 25 or 50 GHz, the fixture's dielectric loss, skin-effect attenuation, and via reflections can easily contribute more loss and ringing than the component being measured. If you simply hand that raw S-parameter file to a system team, you will severely mischaracterize the component, falsely failing a good connector because of a poorly designed board launch.

De-embedding is the mathematical matrix operation that strips away the fixture's electrical parasitics, shifting the VNA's calibration reference planes from the physical cable ends directly to the input and output boundaries of the DUT. When executed properly, it isolates the true, uncorrupted behavior of the component. When executed poorly, it injects non-physical gain spikes, artificial ripple, and causality violations that corrupt channel simulations.

What a measurement actually includes

Calibrating a VNA establishes a reference plane — the place where the instrument believes its measurement starts. A standard calibration puts that plane at the end of the cables, because that is where the calibration standards were connected.

Your device under test is not there. Between the cable end and the structure you care about sit a connector, a launch into the board, and some length of fixture trace. The measurement faithfully reports all of it, as though it were one device.

At 20 GHz those fixture elements can contribute more loss and worse return loss than the device does. Reporting them as part of the device changes the quantity being described, while ignoring them leaves the result without a defensible reference plane. De-embedding moves the reference plane to the structure of interest.

The problem

A VNA measures from its calibration plane, which is the end of its cables. Between there and the structure you care about sit a connector, a launch, and some length of fixture trace. At 20 GHz those contribute more loss and worse return loss than the device under test. Reporting them as part of the device combines fixture and device behavior; ignoring them makes the result difficult to interpret or compare.

Three ways to remove them

TRL — thru, reflect, line. A calibration using fabricated standards on the same board as the device, so the reference plane moves onto the fixture itself. It can be highly accurate because the standards share the device's manufacturing. It costs board area for the coupons and needs several line lengths to cover a broad band.
2x-thru. Fabricate a structure that is exactly two fixture halves back to back, measure it, split it, and subtract one half from each side of the device measurement. Simple, cheap, and the industry workhorse. It assumes the two halves are identical and that splitting the measured thru is valid.
AFR — automatic fixture removal. Algorithmic, deriving the fixture's characteristics from the measurement itself, often with a 2x-thru as input. Convenient and widely implemented; how well it works depends on how well the fixture matches its assumptions.

Checks that expose an invalid result Test the full power-normalised matrix for passivity, then inspect causality with the same extrapolation and window used by the time-domain flow. These checks catch nonphysical results, but passing them does not prove fixture accuracy. Re-embed the extracted DUT between the fixture halves and compare that cascade with the raw measurement; the residual shows what the extraction failed to reconstruct.

Where the error comes from

From confusing three different operations that all get called “de-embedding”.

Subtracting decibels removes magnitude and nothing else. It ignores phase and every reflection, so any mismatch at the interface is simply mis-assigned.
Port extension rotates the reference plane in phase — it moves where you are looking, but removes no loss and no reflection.
Full network de-embedding is the real operation, and it does account for mismatch. Convert each network to its cascade (T) form and the fixture halves divide out: T_DUT = T_left⁻¹ · T_meas · T_right⁻¹. Multiple reflections between fixture and device are handled by the algebra, not assumed away.

So mismatch is not what breaks it. What breaks it is not knowing the fixture halves accurately — the correction can only be as good as the model you divide out, and a poorly characterised launch or an asymmetry between the two halves puts its error straight into the result. The symptom is a ripple in the de-embedded insertion loss with a period set by the fixture length, which is a useful fingerprint because a genuine device rarely produces one.

Two habits make this manageable. Keep the fixture as short and as well-matched as the launch allows, so there is less to remove and less to get wrong. Keep the raw measurement alongside the de-embedded one: when a result is questioned, the ability to show what was actually measured before any arithmetic makes the result reproducible and easier to challenge constructively.

Why this decides how your board behaves

It decides whether a measurement and a simulation are comparable at all. A simulation from the board and a measurement including the launch are answering different questions, and the difference looks like a modelling error.
The error it introduces is frequency-dependent. A de-embedding that is slightly wrong produces ripple, not a constant offset — which reads as a discontinuity in the device that is not there.
Over-de-embedding is possible and produces impossible results. Remove a little too much and the device appears to have gain. That is a useful alarm, because it is unambiguous.
It is where most correlation arguments actually live. When simulation and measurement disagree, the reference plane is a more common culprit than the physics.

What to do about it

Design the fixture at the same time as the device. De-embedding quality is decided at layout: coupon area for standards, identical launches, and enough separation that the fixture and the device are not coupled. Retrofitting good de-embedding onto a board designed without it is not possible.

Check the de-embedded result for physicality. Passivity is the cheapest test there is — a passive device cannot have gain — and it catches over-de-embedding immediately. Causality catches a different class of error.

Close the loop by re-embedding. Cascade the extracted DUT back between the left and right fixture models. Agreement with the raw measurement does not establish a unique DUT, but disagreement directly quantifies an inconsistency in the extraction, port mapping, or fixture models.

Match the simulation's ports to the measurement's reference plane. A plane mismatch is a common source of apparent disagreement, so compare the two definitions before changing the physical model.

Prefer removing a full network to shifting a phase. Port extension only rotates phase; it does not remove loss or the mismatch at the launch. It is quick and it is not de-embedding.

Using this to find a fault

Ripple on a de-embedded S₂₁ that was not in the raw data. The de-embedding is introducing it — usually a mismatch between the assumed fixture and the real one, leaving a reflection behind.
|S₂₁| above 0 dB anywhere. For a passive, consistently power-normalised two-port this is nonphysical. Check normalization, calibration, and fixture removal before concluding which step introduced the excess.
Results that differ between two nominally identical fixtures. The 2x-thru assumption that the halves are identical has failed. Check the launches and the manufacturing.
Good agreement below a few GHz, divergence above. The fixture model is running out of validity, or the standards do not cover the band.

Go deeper — matrix inversion, IEEE 370 2X-Thru mechanics, and non-physical artifacts

The intuitive idea — "measure the fixture, subtract it" — is a useful start, but scalar subtraction omits the phase and mismatch interactions that the full network contains.

Decibels of insertion loss do subtract along a cascade, but only when each stage is matched to the next. The fixture and the device are not matched: there is a mismatch at their junction, energy bounces between it and the connector, and those multiple reflections are part of the raw measurement. Scalar subtraction leaves them in, and they show up as ripple.

Proper de-embedding converts S-parameters into cascading transmission matrices (T-matrices or ABCD parameters). Because cascading networks multiply in T-matrix form ([T_Total] = [T_FixA] · [T_DUT] · [T_FixB]), the device under test is isolated by inverting the fixture matrices: [T_DUT] = [T_FixA]^-1 · [T_Total] · [T_FixB]^-1. Inverting complex matrices removes mismatch interactions and standing waves alongside insertion loss.

The IEEE 370 2X-Thru methodology. Modern multi-gigabit engineering relies heavily on the IEEE 370 standard for de-embedding PCB fixtures up to 50 GHz and beyond. The workhorse method is the 2X-Thru: fabricating a representative calibration trace that represents two identical fixture halves placed back-to-back (Fixture A + Fixture B) without the DUT in the middle.

Time-Domain Splitting: The algorithm measures the 2X-Thru coupon in the frequency domain, transforms it to the time domain via Inverse FFT, identifies the exact physical symmetry midpoint via time-domain reflectometry (TDR), and splits the response into two 1X fixture models.
Impedance Correction: IEEE 370 incorporates time-domain peel-back algorithms to account for localized impedance variations along the lead-in traces, preventing residual launch reflections from leaking into the extracted DUT model.

When 2X-Thru de-embedding fails. The 2X-Thru approach is computationally robust, but its accuracy is strictly bound to physical assumptions:

Fixture Asymmetry: If local PCB etching tolerances, copper thickness variations, or fiber-weave glass bundles cause the left fixture half to differ from the right half, the mathematical split distributes errors unequally, leaving false ripples in the DUT S-parameters.
Mode Conversion: In differential channels, any geometrical asymmetry in the launch vias generates common-mode noise (S_CD21). A standard single-ended or uncoupled de-embedding algorithm cannot account for cross-modal energy transfer.
Near-Field Evanescent Coupling: Matrix cascading assumes networks are separable and interact only via propagating TEM transmission-line modes. If the fixture launch via is placed within a couple of substrate heights of the DUT solder pads, non-propagating reactive fields couple directly between fixture and DUT, invalidating the cascade model entirely.

Diagnosing non-physical artifacts in de-embedded data. Always inspect de-embedded S-parameters before importing them into channel simulators:

Passivity Violations: If insertion loss curves ever rise above 0 dB (|S₂₁| > 1.0) or total scattered energy exceeds unity (Σ |S_ij|² > 1.0), the matrix inversion has over-compensated for fixture loss. In a transient simulator, an un-passive S-parameter model acts as an active amplifier and causes the solver to diverge or oscillate uncontrollably.
Causality Violations: When transformed to the time domain, a physical passive device cannot output energy before the input pulse arrives (t < 0). Truncating high-frequency measurement data or poor phase unwrapping generates unphysical pre-cursor ripples that violate Kramers-Kronig relations.
Periodic Ripple in S₂₁: Regular periodic undulations across frequency indicate un-cancelled standing wave reflections caused by an impedance mismatch between the 2X-Thru calibration coupon and the actual test board launch.

In the real world

De-embedding does not change the raw observation; it changes the network assigned to the DUT. Two competent engineers can report different S-parameters from the same structure when they choose different reference planes or fixture models, so those choices belong with the data.

Which is why any S-parameter file worth using carries its provenance: what was measured, what was removed, by which method, and where the plane ended up. Without that provenance, another engineer cannot reconstruct what the numbers represent.