Skip to content

Device models testing against measurements - #1026

Open
atorkmabrains wants to merge 9 commits into
IHP-GmbH:devfrom
mabrains:device_models_testing
Open

Device models testing against measurements#1026
atorkmabrains wants to merge 9 commits into
IHP-GmbH:devfrom
mabrains:device_models_testing

Conversation

@atorkmabrains

@atorkmabrains atorkmabrains commented Jun 30, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

This PR adds a framework that verifies the SG13G2 ngspice device models against IHP's measurements
(ihp-sg13g2/libs.doc/meas), in ihp-sg13g2/libs.tech/ngspice/testing/devices/, and a CI workflow that runs it.

What is tested

devices measurements what is judged
MOS LV / HV, nmos / pmos (PSP103 through OpenVAF / OSDI) DC transistor sweeps, junction diodes, C-V (junction, gate, oxide) every curve; the V_DS = 0 point of I_D / I_S is not judged (no channel current flows there, the meter reads its offset)
HBT npn13g2, npn13g2l, npn13g2v (VBIC) DC Gummel and output curves; S-parameters: C_jE, C_jC, f_T at 30 GHz as the MQA for SiGe HBTs asks (G. G. Fischer, IHP, 2025-12-16): its characteristics (fg_vcb0, fg_vce, fo_vb on I_C, spar_vb, spar_vc, spar_vcb025 / spar_vcb05), bias windows and structures (npn13g2_T03, npn13g2l_T01, npn13g2v_T01); the other 11 measured structures run on request
pnpMPA (Gummel-Poon) DC of 14 DUTs (forward / reverse Gummel and output), C-V every curve (not covered by the MQA)

How a curve is judged

  • A curve (one measured data block and quantity) passes when the measurement lies inside the band of the
    bcs / wcs corner simulations, widened by corner_tolerance_percent, for all but threshold_percent_oob of
    its points. The typical simulation is judged against the same band.
  • Exceptions (waivers/): a reviewed failing curve keeps the exit code at 0 only while the simulation
    reproduces its recorded failing points exactly (per toolchain fingerprint); every report still counts it as
    a failure. The current entries are recorded from the runs and pending review.

Layout

  • models_verifier/: MDM parsing, DC / S-parameter / C-V runners (ngspice), range checker, exception store and
    generator, reports (final_reports/: results_summary.csv, full_results.csv, new_failures.csv,
    stale_exceptions.csv, final_summary.md)
  • configs/: one YAML per device (and HBT test structure): measurement folder, corners, metrics, tolerances,
    MQA windows; netlist templates
  • waivers/: the committed exceptions
  • tools/: pinned toolchains (toolchain.lock, toolchain.py: ngspice 46 / 47, OpenVAF-Reloaded b4517ad /
    fdf2522), CI helpers, HBT de-embedding check
  • validation/plot_verification_results.py: measured-vs-simulated figures; workflow_notebooks/: notebooks
  • Makefile: make test-all (or test-mos, test-hbt, test-hbt-sparam, test-pnp, test-cv,
    test-<case>), make waive-<case>, make plot-<case>, make unit-tests, make check-deembedding,
    make eda-mos (MOS error report, not a gate)

CI (.github/workflows/device-testing.yml)

Thin: every step calls a make target or a script. Each pinned ngspice version and OpenVAF-Reloaded commit is
built once and cached; then one job per toolchain and make target, plus the unit tests and the de-embedding
check. A job fails on new failures, simulation errors, or exceptions whose curve is missing; its summary lists
the counts.

Run locally

cd ihp-sg13g2/libs.tech/ngspice/testing/devices
python3 tools/toolchain.py list
make test-all TOOLCHAIN=ng47-ovr-b4517ad

Details: README.md in that folder.

@KrzysztofHerman

KrzysztofHerman commented Jul 22, 2026

Copy link
Copy Markdown
Contributor

HI @atorkmabrains, is there any passivity to simplify this PR. It has 53 commits and 901 files changed.

@atorkmabrains
atorkmabrains marked this pull request as draft July 22, 2026 17:38
@atorkmabrains

atorkmabrains commented Jul 22, 2026

Copy link
Copy Markdown
Contributor Author

Hi @KrzysztofHerman

I'll simplify it.

@atorkmabrains
atorkmabrains force-pushed the device_models_testing branch from 1ea8d97 to fd1e6c0 Compare July 22, 2026 18:03
@atorkmabrains atorkmabrains changed the title Device models testing with EDA Device models testing against measurements Jul 22, 2026
@atorkmabrains

atorkmabrains commented Jul 22, 2026

Copy link
Copy Markdown
Contributor Author

@KrzysztofHerman Simplified, but not ready for merge. Please wait.

@atorkmabrains
atorkmabrains marked this pull request as ready for review October 15, 2026 07:39
@atorkmabrains
Add ngspice device model verification framework
a76338d 
Model-vs-measurement verification for 8 devices (4 MOS, 3 HBT, pnpMPA):
MQA bias-window gating with fixed limits, waiver/baseline system,
MOS error EDA (clustering + rule-based labeling + reports), and HBT
S-parameter CjE/CjC/fT extraction and checks.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Add device-testing CI workflow wired to make targets
90b6432 
Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Fix DC bias fidelity and current-sign handling in the device verifier
f94102c 
- Honour IC-CAP SYNC inputs: keep the swept master as the sweep variable,
  carry the slave relation into the netlist and reference the slave source
  to a VCVS copy of the master node. Constant-VCB Gummel sweeps (fg_vcb0,
  fg_vcb05, fg_vcbm05, ...) were simulated with the base grounded, VBE = 0.
- Run sweeps that cross 0 V as two segments from 0 V outward, and keep the
  points of a corner that aborts part-way instead of dropping the whole
  characteristic for every corner.
- Clip currents sign-preserving and score points where measurement and both
  corners sit at the noise floor as agreement. Negative currents (PMOS id,
  NMOS is, pnp ib/ic, HBT reverse ib) were collapsed onto the positive floor
  on both sides and passed vacuously.
- Correct the pnpMPA DUT area table to m^2; values were 1e6 times too large.
- Add design_param_overrides for the DC and S-parameter paths, make the
  template sweep branches case-insensitive, sort input_vars for a
  deterministic column order, add unit checks and a README section.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Simulate npn13g2l with Nx=4 as in the MQA default variant
356e9b2 
The npn13g2l_T01 MDM headers carry REMARKS "Nx=5", while the MQA proposal
defines the default variant as le=2.5 um, Nx=4 and the model card limits
npn13G2l to Nx=1-4. Simulated with Nx=5, measured IC and CjE sit at 0.80
and 0.81 of the typical model, exactly the 4/5 expected from a mislabelled
emitter count; with Nx=4 they sit at 0.92 and 1.007.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Regenerate all waivers on ngspice-47 after the bias-fidelity fixes
5d1090b 
DC waivers for the 8 devices and S-parameter waivers for the 3 HBTs were
rebuilt from scratch with ngspice-47. The previous baseline encoded
simulations at the wrong bias (SYNC sweeps), vacuous passes of clipped
negative currents and the pnpMPA area error rather than model-versus-
measurement differences. All 15 suite steps exit 0 with the new files.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Add plot targets for measured-vs-simulated figures
49e019d 
- Makefile: plot-<device>, plot-hbt, plot-pnp, plot-mos and plot-all wrap
  validation/plot_verification_results.py; figures go to figures/<device>/
  (FIG_DIR=<dir> redirects), HBT S-parameter figures to figures/<device>_sparam/.
- validation/plot_verification_results.py: infer sweep_var, output_vars and
  block_id when a combined_results table lacks them, so the S-parameter tables
  (cje, cjc, ft versus bias) plot the same way as the DC characteristics.
- README: "Generating figures" section; .gitignore: figures/.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Device tests: exact exceptions, pinned toolchains, new coverage, thin CI
c998294 
Exceptions (schema 2): one entry per curve with, per toolchain, the number
of judged and failing points and a fingerprint of the failing points (bias,
simulated tt/ss/ff, 6 significant digits). A failure is waived only when it
is reproduced exactly; reports count waived curves as failures; exit 1 only
on new failures, simulation errors, or exceptions whose curve is missing
from the run. make waive-* records the failures of a finished run as
auto-waived, pending review; every current failure is recorded for the four
toolchains (bipolar devices per ngspice version).

Toolchains: toolchain.lock pins ngspice 46/47 and OpenVAF-Reloaded
b4517ad/fdf2522; tools/toolchain.py builds, registers and compiles PSP103;
make TOOLCHAIN=<id>; every run records toolchain.txt and checks that the
binaries belong to the ID.

Checks: every measured point is kept (the VG = 0 V point of every MOS Id-Vg
curve was dropped before); a point the simulation did not give fails its
curve; unreadable files, unmatched blocks and duplicate curve keys stop the
run. Split DC sweeps restart from a fresh circuit for the second segment.

New coverage: MOS junction diodes (diode template), pnpMPA with 14 DUTs,
11 more HBT test structures (configs/hbt/variants), extra HBT S-parameter
files (self de-embedding, declared copies, de-embedding check), MOS and
pnpMPA C-V tier (cv_runner).

Figures show the judged band and bias windows (shared plotter). Thin GitHub
Actions workflow: cached ngspice / OpenVAF / PSP103 builds, unit tests, one
job per make target and toolchain, job summary with counts, new failures
and the toolchain record.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
Device tests: HBTs checked as the MQA asks, MOS V_DS = 0 point not ju…
ba0cf94 
…dged

- MOS: valid_range on id / is leaves out the V_DS = 0 point of the output
  curves (no channel current flows there; the meter reads its offset).
- HBT: only the MQA ranges are judged (mqa_ranges_only, RangeChecker
  mqa_only); fT only on the MQA file, at 30 GHz; fo_vb on I_C only; only the
  MQA structures (npn13g2_T03, npn13g2l_T01, npn13g2v_T01) are part of
  test-all, waive-all and the CI, the other test structures stay runnable on
  request.
- Exceptions recorded again on the four toolchains; those of curves that are
  no longer checked are removed.
- Tests for the MQA windows, fT at 30 GHz and the HBT configs; README updates.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
@atorkmabrains
HBT config: declare the copied S_deemb of npn13g2_T03 spar_vcb10
d69ad2a 
spar_vcb10's IC-CAP S_deemb is a copy of spar_vcb05's. Neither file is judged
since fT is checked on the MQA file only, so the copy was no longer declared
and the de-embedding check stopped. Declared with sparam_copied_deemb.

Signed-off-by: Amro Tork <amro_tork@mabrains.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@KrzysztofHerman KrzysztofHerman Awaiting requested review from KrzysztofHerman KrzysztofHerman is a code owner
@Mauricio-xx Mauricio-xx Awaiting requested review from Mauricio-xx
@sergeiandreyev sergeiandreyev Awaiting requested review from sergeiandreyev
@dwarning dwarning Awaiting requested review from dwarning
@abdur-rahman-dev abdur-rahman-dev Awaiting requested review from abdur-rahman-dev

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@atorkmabrains @KrzysztofHerman