How does de-embedding work in coefficient extraction?

De-embedding removes the effects of test fixtures, probe pads, and interconnects from the measured data to isolate the device under test (DUT). The most common technique is the cascade (ABCD matrix) method: the test structure is modeled as a cascade of fixture-input, DUT, and fixture-output sections. If the fixture sections are characterized separately (using open, short, thru, and line standards), their ABCD matrices can be inverted and multiplied out from the total measured ABCD matrix to yield the DUT-only matrix. For on-wafer measurements, the open-short de-embedding technique is widely used: measure the DUT structure, then measure the same pads with the DUT removed (open) and with the pads shorted (short). The Y-parameters of the parasitics are subtracted from the DUT measurement. For frequencies above 40 GHz, line-based de-embedding (TRL, multiline TRL) provides better accuracy because lumped-element pad models break down. The de-embedding accuracy directly limits the extraction accuracy, particularly for small devices with parasitics comparable to the fixture effects.

How is extraction accuracy validated?

Validation ensures the extracted model is not overfitting the training data. The primary technique is split validation: extract model parameters from data at one set of conditions (frequency range, bias points, temperatures) and verify predictions at different conditions. For a transistor model extracted from S-parameters at Vds = 10V, validate by predicting S-parameters at Vds = 8V and 12V. If the model is physically based (not just curve fitting), it should interpolate accurately. Error metrics include maximum magnitude error in dB (should be below 0.5 dB for S11/S22, below 0.2 dB for S21), phase error in degrees (below 5 degrees), and stability verification (all poles must have negative real parts for a passive model). Causality and passivity enforcement ensure the extracted model is physically realizable: passivity requires that the S-parameter eigenvalues have magnitude at most 1 at all frequencies, and causality requires that the impulse response is zero for t less than 0. Tools like IDEM (from Politecnico di Torino) and broadband SPICE (from Keysight) automatically enforce these constraints during extraction.

RF Design

Coefficient Extraction

Q: What are the main extraction techniques?

Three families of techniques are used. Direct analytical extraction solves closed-form equations at specific frequencies to obtain circuit element values. For a series RLC circuit, the impedance Z = R + j(wL - 1/wC), so R = Re(Z), L is extracted from the slope of Im(Z) versus frequency, and C from the zero crossing. This is fast but assumes the model topology is exactly correct and works only for simple circuits. Optimization-based fitting defines a parameterized model and adjusts component values to minimize the error between model and measured S-parameters across the full frequency range. Cost functions are typically weighted least-squares of S-parameter magnitude and phase errors. Gradient-based (Levenberg-Marquardt) or evolutionary (genetic algorithm, particle swarm) optimizers search the parameter space. This handles complex models but can converge to local minima. Rational function fitting approximates S-parameters as ratios of polynomials in complex frequency (s = j*omega), producing a state-space or pole-residue model. Vector Fitting is the most popular algorithm, producing broadband macromodels suitable for transient simulation.

/koh-eh-fish-unt ek-strak-shun/

Coefficient extraction derives model parameters (S-parameters, ABCD matrices, equivalent circuit values, or polynomial coefficients) from measured or simulated RF data. Steps: de-embedding (removing fixture/pad effects), data fitting (optimization or analytical), and validation (predicting unseen conditions). Techniques: direct analytical, optimization-based (Levenberg-Marquardt, genetic algorithms), and rational function fitting (Vector Fitting for broadband macromodels).

Category: RF Design

Key step: De-embedding

Accuracy: <0.5 dB, <5°

Understanding Coefficient Extraction

Modern RF design relies on accurate circuit models that predict component behavior across frequency, bias, temperature, and manufacturing variation. These models must be extracted from physical measurements because process variations, parasitic effects, and electromagnetic coupling cannot be predicted purely from design intent. A transistor's small-signal model might have 15 to 30 parameters (R_gs, C_gs, g_m, C_ds, L_s, etc.); extracting these from S-parameter measurements at multiple bias points and frequencies is a non-trivial optimization problem that requires domain expertise, appropriate de-embedding, and careful validation.

The extraction workflow starts with measurement: the DUT is probed on-wafer or in a test fixture, and S-parameters are measured across the frequency range of interest (typically DC to 2 to 3x the operating frequency). The raw measurements include the effects of cables, connectors, probe tips, and pad structures, which must be removed through de-embedding before the DUT's intrinsic parameters can be extracted. The choice of de-embedding technique depends on the frequency range: simple Y-parameter subtraction works below 10 GHz, while TRL or multiline TRL calibration is needed above 40 GHz where distributed effects in the pad structures become significant.

Extraction Equations

De-embedding (ABCD cascade):
A_DUT = A_fix-in^-1 × A_meas × A_fix-out^-1

Cost Function (optimization):
E = ∑_f ∑_i,j w_ij |S_ij,meas(f) - S_ij,model(f)|²

Vector Fitting:
S(s) ≈ ∑_n c_n/(s - a_n) + d + s×e

Where A = ABCD matrix, w_ij = weighting factors, c_n = residues, a_n = poles. Typical target: |S₂₁| error < 0.2 dB, |S₁₁| error < 0.5 dB, phase error < 5°.

Extraction Techniques Comparison

Technique	Model Type	Speed	Accuracy	Best For
Direct analytical	Lumped circuit	Fast	Moderate	Simple RLC, FET intrinsic
Optimization (gradient)	Any parametric	Medium	High	Transistor models, filters
Optimization (evolutionary)	Any parametric	Slow	High	Complex, multi-minima
Vector Fitting	Rational polynomial	Fast	Very high	Broadband macromodels
Neural network	Black-box	Fast (inference)	Variable	Process variation modeling

Common Questions

Frequently Asked Questions

How does de-embedding work?

ABCD cascade: invert fixture matrices and multiply out from total measurement. On-wafer: open-short Y-parameter subtraction (<10 GHz). Above 40 GHz: TRL/multiline TRL for distributed pad effects. De-embedding accuracy directly limits extraction accuracy, especially for small devices with parasitics comparable to fixture effects.

What are the main extraction techniques?

Direct analytical: closed-form at specific frequencies (fast, simple models only). Optimization: gradient-based or evolutionary, minimizes S-parameter error (handles complex models, may have local minima). Vector Fitting: rational polynomial in complex frequency, produces broadband macromodels for transient simulation. Choice depends on model complexity and application.

How is accuracy validated?

Split validation: extract at one condition, predict at others. Error targets: |S₂₁| < 0.2 dB, |S₁₁| < 0.5 dB, phase < 5°. Stability: all poles must have negative real parts. Passivity: S-parameter eigenvalues ≤ 1 at all frequencies. Causality: zero impulse response for t < 0. Tools auto-enforce these constraints.

Related Terms

← Codebook CSI Coefficient Of Thermal Expansion →