Skip to content
Hawse1337Public

About

An antenna design bench: parameter studies, guarded AEDT solves, artifacts kept, targets changeable. Python library + agent skill; web GUI excluded.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

aster — an antenna design bench, evidence first

tests release docs license: MIT python 3.10+

Keep the measurement. Change the target.

aster was previously published as antopt. The package import, CLI and environment prefix are now aster / ASTER_*; the old repository URL redirects.

aster is a design bench for antenna studies. You describe a parameter space and what "good" means, the library drives candidates through a guarded AEDT session, and every solve leaves its raw artifacts behind. When the requirement changes — a different band, a tighter ripple, a flatter elevation cut — you re-judge the history instead of burning another night of solver time.

中文简介:README.zh-CN.md
The web GUI / desktop application is under validation and is not part of this repository.

The five-station loop

Three forms, one bench

form how you get it use it when
Python library + CLI wheel/sdist from Releases, then aster doctor you script your own loop, run CI, or work in a notebook
Agent skill aster skill install (or copy skills/aster/) an AI agent should run design studies with evidence rules
Docs + releases hawse1337.github.io/aster you want the contract, the workflow and pinned artifacts

The bench

Four verbs, one loop:

  • Design — ParamSpec / ParamSpace describe the knobs, Objective / Constraint describe the target. Nothing here knows about AEDT.
  • Solve — a candidate becomes a real run through the companion guard (ansys-aedt-mcp). Sessions are addressed explicitly; budgets are spent, not assumed.
  • Keep — S11 Touchstone, far-field grids and native report images stay with the run record. A number without its artifact is a rumour.
  • Judge — acceptance rules are code (band_acceptance, pattern_acceptance, acceptance), so a verdict can be recomputed, compared and challenged.

The archive is the point. A campaign is just a Python recipe over that loop; swap the evaluator (dry model or real solve) and the same recipe still runs.

What it refuses to hide

risk how aster handles it
"the result looks great" every measured run keeps result.s1p, farfield_views.json and the native report PNGs
"we changed the target, re-run everything" records are re-scored; the solver is not called again unless you ask
"the surrogate says it is good" dry/surrogate results are labelled PREDICTED; validation state is explicit
"the solve died, we lost the point" failed runs are recorded as failed, with the reason, not dropped
"which AEDT session was that?" the guard requires one explicit PID or gRPC port per call

Try it in one minute

py -3.10 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -U pip
.\.venv\Scripts\python.exe -m pip install -e ".[optimizer,surrogate,plot]"

aster version
aster doctor              # what is missing before a real campaign
aster dry-demo --samples 64

dry-demo is a complete study on an analytic objective — Latin-hypercube design, Pareto archive, no AEDT required. The same shape in code:

from aster.core import Objective, ParamSpace, ParamSpec, latin_hypercube
from aster.optimize import Evaluation, pareto_archive

space = ParamSpace([
    ParamSpec("ring_radius", "mm", 20.0, 60.0, 40.0),
    ParamSpec("ring_height", "mm",  1.0, 10.0,  4.0),
])
objectives = [Objective("gain", "peak_gain", "maximize"),
              Objective("bw", "s11_bandwidth", "maximize")]

runs = [Evaluation("c%02d" % i, "ring", p,
                   metrics={"peak_gain": predict_gain(p),
                            "s11_bandwidth": predict_bw(p)}, feasible=True)
        for i, p in enumerate(latin_hypercube(space, 32, seed=1))]
front = pareto_archive(runs, objectives)

Runtime safety, by construction

Real solves go through the companion guard ansys-aedt-mcp, not through a loose PyAEDT call in the same process:

  • the guard records the AEDT sessions it launched and refuses to attach to unmanaged ones (guard_status classifies every live ansysedt.exe);
  • run_python_code / run_python_script are refused by default at the guard's MCP and worker layers; a human operator can enable them for a supervised session;
  • run tokens bind analysis submissions to project/design/setup when AEDT_GUARD_ENFORCE=1;
  • launch/release events are audited and PyAEDT app objects are checked against the verified desktop.

If the guard MCP tools are missing, stop — do not fall back to shell-level PyAEDT or ansysedt.exe. Fail-closed DSH/Codex configurations and the threat model are in the guard repo. See user_doc/safety.md for the boundary and residual risks.

Real studies

  1. Install the guard system and its virtual environment (PyAEDT 1.10.0, AEDT 2026 R1).
  2. Point ASTER_AEDT_EXE, ASTER_GUARD_REPO, ASTER_VENV_PYTHON, ASTER_REFERENCE_PROJECT, ASTER_CAMPAIGN_ROOT at your machine.
  3. Run the worker against a campaign directory; each candidate gets its own run folder and archive row.

The contract, the run lifecycle and the artifact layout are in user_doc/aedt_integration.md, with a walkthrough in examples/aedt_campaign.md.

Drop it into an agent

The skill lives in skills/aster/: SKILL.md plus references/ (API map, AEDT contract, evidence rules) and a runnable scripts/dry_demo.py. Install it with one command:

aster skill install            # → %USERPROFILE%\.dsh\skills\aster
aster skill path               # print the destination

Or copy skills/aster/ into your agent's skill directory by hand. The skill does not cover the GUI, which is not published.

Repository map

assets/        banner + loop illustrations
aster/        the library
  core.py        parameter spaces, objectives, constraints
  optimize.py    evaluations, dominance, Pareto archive
  campaign.py    campaign recipes
  runtime/       measured-run manager + job store (no web server)
  worker.py      command-line measured campaign worker
  aedt_host.py   AEDT process helpers
  skills/aster/  bundled skill (mirrored from skills/aster at build time)
  cases/  benchmark/  interop/  optimizer/  parametric/  physics/  surrogate/  topology/
skills/aster/  canonical agent skill (SKILL.md + references + scripts)
tools/         sync_skill_bundle.py (mirror the skill into the package)
user_doc/      installation, architecture, workflow, AEDT integration, skill, validation
examples/      dry study + measured-campaign walkthrough
tests/         offline suite (no AEDT needed)
mkdocs.yml     documentation site

Status

  • Offline suite: 782 passed, 80 subtests; CI runs it on Windows / CPython 3.12.
  • Measured runs are exercised on AEDT 2026 R1 through the guard; any solver number quoted from a study must be accompanied by its run artifacts.
  • Wheel and sdist: v1.54.0 release; documentation: hawse1337.github.io/aster.

License

MIT — see LICENSE. If you build a study on top of this bench, keep the artifacts; a bench without evidence is just furniture.

About

An antenna design bench: parameter studies, guarded AEDT solves, artifacts kept, targets changeable. Python library + agent skill; web GUI excluded.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages