dynmodels_examples

Dynamic (AC / RF) compact-model validation — Enhancement-161

Enhancements 159 and 160 validated the CMC compact models' DC behavior. This one exercises their dynamic behavior — the part that matters most for analog and RF — which flows through a completely different code path: OSDI's reactive (charge) Jacobian stamping and ngspice's .ac analysis. The models are the CMC reference decks bundled with OpenVAF, compiled in place.

BSIM4 C-V — gate capacitance vs bias

The gate capacitance Cgg(Vgs) is extracted from .ac (Cgg = Im(I_gate)/ω) and swept over gate bias. It rises from a small subthreshold value (~40 fF, overlap + fringe) to the oxide capacitance in inversion (~129 fF) — the textbook MOSFET C-V curve — and the OSDI model matches ngspice's built-in BSIM4 to < 1 % at every bias. Because the capacitance is dominated by the (version-independent) oxide term, this is an even tighter check of the reactive stamping than the DC I-V match.

Cutoff frequency fT — AC current gain roll-off

The cutoff frequency fT is where the AC current gain |h21| = |I_out/I_in| falls to 1. Panel B shows the -20 dB/decade roll-off and the |h21|=1 crossing:

• BSIM4 (MOSFET) — fT ≈ 3.5 GHz; OSDI matches the built-in to ~1 %.
• HICUML2 (SiGe HBT) — ngspice has no built-in bipolar reference. The default model has zero transit time (t0=0) → infinite fT, so a realistic dynamic parameter set is supplied (t0=10 ps, 1 fF junction caps). The resulting fT sits right at the transit-time limit 1/(2π·t0) ≈ 15.9 GHz and rises with collector current — textbook bipolar behavior.

Run it

python3 verify_dynmodels.py    # 4 checks, under BOTH the Sparse and KLU solvers
python3 make_dynmodels_fig.py  # -> dynmodels_ac.png

Why the results are physically correct

• C-V shape. In subthreshold the gate sees only overlap/fringe capacitance; as the channel inverts, the gate couples to the channel through the oxide, so Cgg climbs toward Cox·W·L. Matching the built-in model to sub-percent confirms the OSDI charge model and its reactive Jacobian are stamped correctly.
• fT. A single-pole roll-off gives |h21| ≈ fT/f, so |h21|=1 at f=fT. For the MOSFET fT ≈ gm/(2π·Cgg); for the bipolar, charge-control theory gives fT ≈ 1/(2π·τ_f) in the transit-time-limited regime — exactly where the measured 15.5 GHz lands for t0 = 10 ps.

Notes

• .ac is supported under both the Sparse and KLU solvers, so every check runs under both.
• This is a validation/example enhancement — it exercises the existing toolchain and needs no ngspice/openvaf-r source change.

See Enhancement-161 for the full write-up.