Skip to content
geer1895Public

About

RF/microwave design automation + LLM orchestration: 16 EM/EDA engines behind one interface, 75 parameterized template families, 340 deterministic calculators, 87 calibration anchors; AI agents drive everything via MCP (157 tools) — physics numbers come only from deterministic kernels, never from the LLM. 28000+ unit tests, GPL-3.0

Topics

Resources

Code of conduct

Contributing

Stars

47 stars

Watchers

1 watching

Forks

Latest commit

 

History

143 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

rfauto — RF/microwave design automation with LLM orchestration

rfauto

tests Python License code style: ruff

English | 简体中文

rfauto turns RF/microwave design from "draw, wait for the solver, copy numbers" into one reproducible automated loop. Describe the device you want; it derives first-cut dimensions from closed-form physics, drives whichever solver you have, runs every result through quality gates that catch numerical artifacts, then optimizes the dimensions — and hands you a report where every number carries its provenance. AI assistants can drive all of it over MCP, under one iron rule:

Every physical number (frequency, loss, geometry) is produced by the deterministic kernel or by a solver — never invented by the language model.

Measured counts for this revision (exact; verifiable with scripts/check_numbers.py):

CALCULATOR_REGISTRY 340 (with experimental 341) · TEMPLATE_META 76 · EXPECTED_TEMPLATES 76 · ANCHORS 87 · CLI commands 357 (leaves) · (208 MCP tools + 17 resources)

Scale numbers: MCP tools 208 (+17 resources) | CLI commands 357 (leaves) | CALCULATOR_REGISTRY 340 [with experimental 341] | device templates 76 | anchors 87

Contents · Capability map · Quick start · Why trust it · Engine ladder · Architecture · Docs map · Roadmap

Capability map


1. Capability map (nine domains)

Each domain works on its own; together they form the loop above. Every command below is copy-pasteable.

1.1 Device modeling and synthesis

76 parameterized device templates — dividers, couplers, filters, antennas, transitions, lumped IPD passives. Each template synthesizes first-cut dimensions from closed-form physics and registers the acceptance criteria for that family, so you can tell a real result from a grid artifact. Line widths are never copied from a document: they come from skrf Hammerstad-Jensen synthesis (single source).

rfauto syn mline 50.0 --freq 2.4 --stackup rogers4350b_h0.508
#  width: 1.1133 mm    eps_eff: 2.8530
rfauto templates list                  # the 75 families and their parameters
rfauto templates golden-diff           # geometry golden regression

A recipe is the input: one YAML file describes frequency, substrate, targets and constraints, and every channel reads the same file.

1.2 Multi-engine simulation

One adapter layer, 16 channels: HFSS, ADS, openEMS, COMSOL, Elmer, NGSolve, Meep, Icepak, Q3D, Palace, KiCad, ngspice, FDTDX (JAX), MMT (analytical mode-matching), QucsatorRF, and a VNA measurement channel. Commercial engines are optional extras; the whole core runs on the built-in fake channel — no license needed to try everything.

rfauto solvers list                    # registered channels and availability
rfauto run recipes/patch_antenna_v1.yaml --adapter fake

1.3 Optimization and tuning

TPE / CMA-ES / NSGA-II for multi-objective work (Pareto fronts), plus a surrogate route: fit a cheap model on a batch of results, search on the model, and spend expensive solves only where they matter. Batch campaigns carry budget admission, quotas and watchdogs, so they can run unattended overnight.

Tuning (real run: one row per trial, the curve is the best-so-far cost)

rfauto tune  recipes/wilkinson_pd_v1.yaml --adapter fake --max-trials 40
rfauto sweep recipes/wilkinson_pd_v1.yaml --adapter fake
rfauto campaign plan --config campaign.yaml

1.4 Quality gates

Solvers fail quietly: a coarse mesh shifts the resonance, a wrong port solves to ±0 dB, a truncated pulse pushes |S11| above 1 — and the garbage looks as plausible as the truth. rfauto makes the judgment an explicit gate.

Quality gates

Six gates cover six classes of failure: port semantics (five geometric criteria at model-build time), grid artifacts (energy / convergence / minimum spacing), physics invariants (passivity / reciprocity / symmetry), cross-engine arbitration (a second engine on the same structure), artifact freshness (manifest fingerprint vs files on disk), and anchor drift. A result that misses a gate is reported FAIL with the gate name and the margin — never silently passed.

1.5 Calibration and anchors

87 calibration anchors cover 68 device families (75 templates ship; families without an anchor say so). Layered by evidence strength: 18 are backed by an HFSS ↔ openEMS dual-engine comparison, 2 by HFSS/MMT arbitration, 1 by HFSS self-consistency; 62 are closed-form or quasi-static FD reference anchors (59 closed-form + 3 FD, each with formula source and validity domain on file); the remaining 4 have no engine comparison and say so — a closed-form anchor is never dressed up as engine-backed. Artifact freshness has four states (fresh/stale/legacy/error): an old file replacing a new run is flagged, and rfauto runs verify exits 1.

rfauto anchors list                    # values / status / provenance
rfauto anchors validate                # check provenance records one by one
rfauto runs verify <run_id>            # freshness (stale ⇒ exit code 1)

1.6 Datasets and the lake

Every run drops a self-describing directory: environment fingerprint, recipe snapshot, metrics, figures, Touchstone files, and a manifest recording hashes of all artifacts. Datasets assemble runs into a queryable lake with tiered compaction and batch collection.

runs/20261026_132409_ff7c0bde/
├── meta.json              # environment fingerprint (versions, git sha, deps)
├── recipe.snapshot.yaml   # the recipe snapshot this run used (replay reads it)
├── manifest.json          # SHA256 + size of every artifact (freshness basis)
├── report.md              # human-readable report: metrics, gates, verdict
└── results/
    ├── metrics.json       # machine-readable metrics
    ├── params.s3p         # Touchstone (real-engine channels only)
    └── figs/              # figures
rfauto runs list
rfauto dataset materialize --help
rfauto lake index

1.7 Remote execution and batch

A machine registry with liveness probes and per-machine mutex: send the heavy work to a compute server and keep only orchestration locally. Remote HFSS runs over gRPC attach; batch jobs launch short-lived processes. Everything is cleaned up on completion (processes killed, scheduled tasks removed, artifacts pulled back) so no orphan holds a license seat.

rfauto remote probe                    # liveness and version check
rfauto remote status

1.8 AI orchestration (MCP)

208 MCP tools and 17 resources: Claude Desktop, Cursor or your own agent can operate the framework. Agent changes land in a sandbox draft and must pass validation gates before touching the workspace; the agent emits typed tool calls only, and the kernel fills in the physics.

pip install -e ".[mcp]"
python -m rfauto.mcp_server            # stdio transport

1.9 Workbench

rfauto ui starts a local review workbench (your data never leaves the machine): 24 views over run history, S-parameters, 3D fields, the template gallery, campaign boards, calibration, datasets, reports and a microwave toolbox. Command palette (Ctrl+K), deep links, whole-site zh/EN switch.

Workbench dashboard

Template gallery (75 families, each card a plain-language parameter guide)

Microwave toolbox (pick a key, fill parameters, read the result — every key carries its formula source)


Quick start

The loop in five steps: specify → design → solve → judge → verified result

Requirements: Python 3.10–3.12 (python --version). No commercial tool is needed — steps 0–6 run entirely license-free in about ten minutes.

Step 0 · Get the code

# A. with git (recommended, easy to update)
git clone https://github.com/geer1895/rfauto && cd rfauto

# B. without git: use Code → Download ZIP on GitHub, or grab
#    "Source code (zip)" from the Releases page and unzip it

Step 1 · Install

python -m venv .venv
.venv\Scripts\activate                  # Linux/macOS: source .venv/bin/activate
pip install -e ".[ui]"                  # core + browser workbench
rfauto --help                           # a command table means it works

Steps 2–6 only need the core (pip install -e .); rfauto ui needs the [ui] extra (starlette/uvicorn); the test suite needs [dev].

Step 2 · Health check

rfauto doctor

Returns a table in a second: Python version, AEDT/ADS detection (marked "not configured" if absent — the license-free flow is unaffected), cache mode. Add --janitor --min-free-gb 8 to also list orphan solver processes and enforce a memory floor — worth doing before long real-engine runs.

Step 3 · First simulation (no license)

rfauto run recipes/wilkinson_pd_v1.yaml --adapter fake

Measured output on the development machine (2.4 GHz Wilkinson divider, Rogers 4350B 0.508 mm):

✓ done  run_id: 20261026_132409_ff7c0bde
  metrics:
    s11_db_max_in_band: -12.4325
    s11_db_min_in_band: -16.2978
    s21_db_mean_in_band: -3.5887
    iso_s23_db_min_in_band: 27.3788

The fake channel is a closed-form analytical model, not an EM solve — its job is to exercise "recipe → run → artifacts → gates" in seconds. Real physics starts at step 7, with the same recipe and the same gates.

Step 4 · Read the artifacts

Each run lands in runs/<timestamp>_<hash>/ (layout in 1.6). Look at report.md (human) and results/metrics.json (machine), then verify freshness:

rfauto runs verify 20261026_132409_ff7c0bde

fresh means the files on disk still match the closing fingerprint; stale (an old file replacing a new one) exits 1 and is never silently passed.

Run detail (S-parameters, 3D model, artifact list)

Step 5 · Open the workbench

rfauto ui                 # http://127.0.0.1:8642, loopback only
rfauto ui --port 8643     # use another port if a stale instance holds 8642

The left rail is the workspace: Dashboard, Run History, S-Parameters, Template Gallery, Microwave Tools, Calibration, Datasets… The zh/EN switch sits at the top right.

Step 6 · Tune one round

rfauto tune recipes/wilkinson_pd_v1.yaml --adapter fake --max-trials 40

Measured: 40 trials in about 46 s, cost dropping from 3.64 to 2.55, with the recommended λ/4 arm 18.00 mm, 70.7 Ω arm width 0.448 mm and 50 Ω feed 1.151 mm (s11_db_max_in_band improves from −11.46 to −12.45 dB). The Tuning view in the workbench lets you watch it trial by trial.

Step 7 · Switch to a real engine

pip install -e ".[hfss]"        # pyaedt; requires AEDT 2025 R1
# or build openEMS (free) following docs/openems_build_guide.md

Replace --adapter fake with --adapter openems or --adapter hfss — the recipe and the gates do not change. Before unattended HFSS runs, read docs/unattended_hfss_automation.md, the operational discipline distilled from a year of real-machine work. The three that bite first:

  • one Hfss instance per Desktop — a second one starts returning None from the design-variable manager;
  • one failed assignment poisons the whole session (the gRPC channel breaks and release_desktop fails silently) — rebuild the session, never continue on a poisoned one;
  • project / outdir / log paths must be absolute — a relative -outdir makes Project.Rename hang for minutes before failing.

Step 8 · Command cheat sheet

rfauto run <recipe> --adapter fake|openems|hfss   # one run
rfauto tune <recipe> --max-trials 60              # outer optimization loop
rfauto sweep <recipe>                             # parameter sweep
rfauto replay <run_id>                            # reproduce a historical run
rfauto templates list | gallery | golden-diff     # templates and geometric goldens
rfauto anchors list | validate                    # anchor registry
rfauto runs list | verify <run_id>                # runs and freshness
rfauto em doctor | diff | metrics | sweep         # EM semantics layer
rfauto ui                                         # workbench

First-run troubleshooting

Symptom Likely cause Fix
rfauto ui won't open, or pages 404 port 8642 held by an old instance (old code) rfauto ui --port 8643; or check the command line and remove the old instance
rfauto: command not found venv not activated, or install failed re-activate; verify with rfauto --help after pip install -e .
AEDT fails to write files (STEP/export) non-ASCII characters in the path use ASCII project/work directories; stage into an ASCII temp dir and copy back
HFSS cannot connect / license errors orphan ansysedt holding a seat rfauto doctor --janitor lists them; review, then --janitor-clean
solve killed midway, machine sluggish not enough free memory rfauto doctor --min-free-gb 8 --enforce before launching
fake results differ a lot from hardware fake is a closed-form model, not an EM solve fake only validates the pipeline; switch to openEMS/HFSS — same recipe, same gates
no Touchstone in the run directory the fake channel only emits analytical metrics read results/metrics.json; Touchstone comes from real engines
Touchstone on disk disagrees with the run record an old file replaced the new one rfauto runs verify <run_id> — stale exits 1

3. Why these results can be trusted

Every result carries provenance: which solver ran it, how the geometry was derived, which gates it passed.

Trust layer

Three layers stack up:

  1. Numbers come only from the kernel. Frequencies, losses and dimensions are either closed-form outputs or solver outputs, each with its formula source and validity domain; the LLM only orchestrates and explains.
  2. Results pass gates before they are trusted. Six gates (see 1.4) cover everything from model-build to post-run; a miss is an honest FAIL with the margin.
  3. Every change leaves a trace. Agent edits go sandbox draft → validation gates → promotion, and the promotion records who proposed it, which gates passed and the exact diff; anchor changes need a signature.

When two engines agree on the same physical structure, the free one can be trusted for sweeps; beyond the pre-declared gate (|Δ| ≤ 0.5 dB and the same direction ⇒ AGREE) it is recorded honestly as DISAGREE — never rounded to green.


4. Engine ladder and arbitration

Engine ladder

Engine License Typical role
HFSS (Ansys AEDT) commercial full-wave reference / arbiter
ADS (Keysight) commercial circuit and system co-simulation
openEMS open source (GPL, subprocess) fast FDTD batch solving
COMSOL commercial FEM multiphysics
Elmer open source multiphysics FEM
NGSolve open source frequency-domain FEM
Meep open source FDTD
Icepak / Q3D (Ansys) commercial thermal / parasitic extraction
Palace open source parallel FEM
KiCad open source PCB DRC and layout extraction (subprocess)
ngspice / Xyce / QucsatorRF open source circuit simulation
FDTDX (JAX) open source differentiable FDTD
MMT open source analytical mode matching
VNA — measurement channel (sim-to-measurement correlation)
fake (built in) project-owned license-free pipeline checks / CI / teaching

Commercial tools require your own valid license; the framework neither bundles nor circumvents any license and ships no vendor-proprietary content (see THIRD_PARTY_NOTICES.md).


5. Architecture

Architecture

Dependencies only point downward and import-linter enforces it in the test gate (a reverse import turns the build red):

cli / mcp_server / ui  →  service  →  linkage / optimization / models
                                      →  adapters  →  pipeline  →  infra  →  core
  • The service layer speaks JSON: features start as service functions; CLI, MCP and UI are thin shells;
  • Numbers live in the deterministic kernel (core/, pure functions): the LLM or an agent never produces a physical number;
  • Replaceable parts are base class + registry: EMSolverAdapter / EMSolverRegistry, CALCULATOR_REGISTRY, the device-template registry.

Repository layout

src/rfauto/
├── cli/            # command line (357 leaf commands)
├── mcp_server/     # MCP server (208 tools + 17 resources)
├── ui/             # browser workbench (24 views, zero-build native ESM)
├── service/        # business layer: JSON in, JSON out
├── linkage/        # HFSS ↔ circuit linkage
├── optimization/   # TPE / CMA-ES / NSGA-II / surrogates
├── models/         # recipes and device models
├── adapters/       # 16 engine/tool channels + template renderers
├── pipeline/       # run pipeline
├── infra/          # infrastructure (runs paths / cache / logging)
└── core/           # deterministic kernel: calculator registry + physics criteria

Calculators by domain

Calculators by domain


6. Docs map

From zero

By topic

Reference

Explanation and operations

Container: a Dockerfile ships at the repo root — docker build -t rfauto:local . && docker run --rm rfauto:local smokes it; the image contains only the open-source core, and mounting your workspace is enough to run for real (-v "$PWD:/work" -w /work).


7. Status and roadmap

rfauto is a working tool, not a demo: the core loop (template → synthesis → solve → gates → optimize → report) has run on real HFSS, ADS, openEMS, COMSOL and KiCad installations, backed by a test suite of the size above on both Windows (full) and Linux CI. Development is Windows-first and single-maintainer; Linux support is improving.

Planned next:

  • Datasets and benchmarks — the simulation datasets and agent evaluation sets accumulated by the data-factory pipeline are not in this repository yet; we plan to publish them progressively and welcome co-maintainers;
  • A methodology paper — on quality gates and the deterministic kernel; co-authors welcome;
  • More device families, more engines, a smoother first run — all good first issues.

If any of that interests you, open an issue — the aim is a community project, not one person's archive.

Contributing

Issues and PRs are welcome — see CONTRIBUTING.md for how to get started and what the #NNN markers in code comments mean.

Citation

If rfauto helps your research, please cite it — see CITATION.cff.

License

rfauto is GPL-3.0 (see LICENSE). Third-party dependency licenses are listed in THIRD_PARTY_NOTICES.md. The optional openEMS adapter drives GPL openEMS in a separate process; this repository does not bundle the openEMS bindings — build them from the official openEMS sources.

Changelog

v0.14.0 (2026-11-10)

Added:

  • EDA/IC parser suite (~30 formats): structured, schema-stamped parsers for SPICE, Liberty, CITI, SDC, IBIS, IBIS-AMI, Verilog netlists, SPECTRE, Verilog-A, GDSII, LEF/DEF, SPEF, SDF, Gerber, Excellon, IPC-D-356, Touchstone, VCD, STIL, SKILL/OCEAN, BSDL, ODB++, STDF, DRC rule files and LVS comparison - each with a dedicated CLI domain and read-only MCP tools
  • Animation pipeline: 17 deterministic chart types (S-parameter evolution, Smith trajectory, tuning race, field heatmap, VSWR history, constellation/eye, budget funnel, ...) with a self-check loop and rfauto anim render CLI/MCP tools
  • EM benchmark suite: deterministic-judge tasks (closed-form arbitration, anchor re-verification, prompt-variant campaigns) with hidden/visible check layers, negative-variant probes, task admission gates (no-op must fail, oracle stability) and out-of-distribution envelope flags; analog-side families (spec-to-VA, testbench generation, end-to-end, bugfix) and inverse-design scoring
  • Neural surrogates: MLP training with leave-one-out evaluation (optional torch extra with numpy fallback), batch evaluation with top-k ranking over LHS candidates, and an active-learning loop that nominates the most-uncertain points for full-wave follow-up
  • AEDT project migration: rfauto aedt-migrate walks legacy projects through a five-stage decision tree (direct open -> health check -> baseline comparison -> bridge -> CAD re-export) with honest verdicts and a per-project report
  • gm/ID lookup sizing kernel: table build, lookup / size / size-from-fT APIs, fingerprint caching and clamped out-of-range flags
  • Session capture: record a working session and compile it into a recipe draft inside the sandbox (production recipes stay untouched)
  • Toolchain lock: capture the local solver fingerprint (HFSS/ADS/ openEMS/python packages) and diff a remote machine against it
  • MCP surface grows to 208 read-only tools and 17 resources: parser tools, SKILL package resources and workflow prompts; every tool now carries annotations (read-only/destructive hints verified by static analysis) and launch-class tools get sliding-window rate limits with a merge-calls hint; stdio channel purity is enforced at runtime
  • Request receipts: idempotency keys with four outcome states (not_started / completed / failed / unknown) for remote and async runs - unknown outcomes are never silently replayed
  • Strict result accessors: read_touchstone(strict=True) validates the read boundary (unique file match, path-escape and symlink escape rejection, ULP-level duplicate frequencies, exact key lookup without alias guessing)
  • veriloga-authoring skill package: teaches agents to write .va models the built-in parser accepts (all examples verified in tests)
  • KiCad schematic generation with ERC parsing and a round-trip certificate; vendor documentation retrieval (SQLite FTS5 over PDF/CHM/HTML with a zero-ingest audit)
  • Engine identity: runs can carry a machine-readable engine/version block; netlist parsing gains an optional semantic overlay (source / confidence / evidence annotations with a coverage gate)

Fixed and improved:

  • CLI grows to 357 leaves: 16 new parser domains wired in one batch with help panels, registration-order snapshots and count anchors re-pinned in the same commit
  • The fake branchline model upgrades to an ideal TL-matrix closed form (Pozar even/odd half-circuit) - benchmark verdicts flip from as-recorded FAIL to PASS with machine-precision physics checks
  • SPEF parser accepts three-value (min:typ:max) entries, scaled products and line continuations with bit-exact recovery pins; SPEF errors now derive from ValueError
  • Schema keys unified across the parser suite ("schema": "domain/1.0") with n_warnings stats completed; verilog keeps a dual-key compat release
  • Exit-code verdicts centralized: a zero exit code with fatal log markers now fails loudly across the SPICE/ADS/openEMS adapters
  • The layered-stack template family closes its root-cause chain (feed-wall structure fix + window-dominance proof); the branchline/ hairpin knowledge notes extend accordingly
  • A recorded-fixture test tier lands: cleaned, manifest-stamped real-machine artifacts replay offline (new recorded marker next to unit / real_edt)
  • Docs: parser-cluster user guide (zh/en) plus an integration guide for the whole suite; docs_site regenerates in the same commit

Current version: v0.14.0 (previous: v0.13.0).

v0.13.0 (2026-11-02)

Added:

  • Pin-level schematics: rfauto diagram schematic renders ADS netlists as schematics with pins, wires and net names
  • SVG export: rfauto diagram to-svg converts .drawio to SVG so diagrams can be viewed on servers and in reports without draw.io
  • 3D stackup preview: rfauto diagram stackup --preview renders a 3D preview PNG alongside the cross-section view
  • Review annotations: annotations from multiple reviewers merge into one file, with per-reviewer colors, attribution and counts
  • Test-setup diagrams: new CISPR 25 conducted-emission scene, with optional professional electrical component shapes
  • ADS tuning: active-chain and HB load-pull templates now support diagram-driven tuning; back-annotation can embed S-parameter curve plots; the optimizer can call ADS circuit simulation as its objective
  • Design reports: rfauto design-report gains Pareto-front and two-run parameter/metric comparison sections
  • Calibration design: rfauto cal-wizard trl proposes Thru/Reflect/Line dimensions, bandwidth coverage checks and a test-setup diagram for a given frequency band
  • Operations: rfauto remote board shows the server and local solve queues; one-command performance profiling; affected tests pre-run automatically when source files change (devtest_speculative --watch)
  • Surrogates: Liquid Neural Network (CfC), complex stamp family (ngo_stamp_rf), multi-port parametric VF, SVGP, per-parameter SPICE netlist export, ONNX export
  • Optimization: TuRBO multi-restart joins the competition leaderboard; SAAS posterior supports NUTS sampling; multi-objective optimization supports constraints (c-EHVI)
  • Architecture credit: event-driven execution, liveness-probe and channel-separation ideas absorbed from DeepSeek Super-agent (MIT) land as speculative test pre-runs, rfauto watchdog and unified CLI/JSON output rules; two design notes in docs/explanation/agent-orchestration.md and docs/explanation/cli-json-channel.md

Fixed and improved:

  • The serial test gate pins BLAS to a single thread, removing flaky timeouts in multi-thread environments
  • A failed comparison gate no longer reports "ADS stage failed"; the two failure kinds now have distinct messages
  • 357 CLI commands, 208 MCP tools, 36454 unit tests (same source as the README badge)

About

RF/microwave design automation + LLM orchestration: 16 EM/EDA engines behind one interface, 75 parameterized template families, 340 deterministic calculators, 87 calibration anchors; AI agents drive everything via MCP (157 tools) — physics numbers come only from deterministic kernels, never from the LLM. 28000+ unit tests, GPL-3.0

Topics

Resources

Code of conduct

Contributing

Stars

47 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages