Skip to content

agents.md

Latest commit

 

History

History
109 lines (86 loc) · 8.31 KB

agents.md

File metadata and controls

109 lines (86 loc) · 8.31 KB
Raw

For AI agents

This page is for an AI agent (or a person scripting one) that drives UltrafastEngine: writing system files and run specs, running them, reading results, and changing the code. The short version is in AGENTS.md at the repository root and llms.txt (an index of the documentation).

The model in one paragraph

A system file (.ufs, JSON with comments) lists elements in beam order (or, in a Branched system, with the link next from each element): bulk, gap, plate, mirror, lens, gratingPair, crystal, gain, ... (supported elements) A run spec (run.json) gives the input pulse, physics switches, instruments and overrides. The engine compiles both into linear and nonlinear operators on a frequency grid that it sizes itself, propagates the pulse, and writes summary.json (scalars), result.npz (arrays), CSVs and provenance.json (where every number came from). Every option has a documented default; nothing is guessed silently -- unknown keys produce warnings with suggestions.

Discover before you write

question ask the engine
every element type, option, unit, default docs/reference/options.md, or ultrafast schema system / ultrafast schema run (JSON Schema)
what is built in (materials, crystals, models, element types, instruments) ultrafast info (JSON; Python uf.capabilities())
materials and refractiveindex.info pages ultrafast materials list, ultrafast materials show <name>
is my file valid ultrafast validate system.ufs run.json (exit code ≠ 0 on errors)
a commented starting point ultrafast template run.json; system files: the examples
what the engine will do with each element ultrafast describe system.ufs
Python signatures docs/reference/python-api.md; TypedDicts in ultrafast.types

Set ULTRAFAST_STRICT_OPTIONS=1 to turn unknown-key warnings into errors while you are writing files: a typo then fails instead of being ignored.

Run and read

ultrafast run system.ufs run.json -o results      # exit 0 on success; warnings on stderr and in summary.json

results/summary.json:

  • output: the final pulse (energyJ, fwhmS, peakPowerW, transformLimitFwhmS, spectralFwhmM, pulseGddS2, ...);
  • elements[]: per element id, bIntegralRad, energyOutJ, peakFluenceJPerM2, lidtMargin, criticalPowerRatio, notes, values (e.g. trackedBeamRadiusXM, peakElectronDensityM3);
  • instruments[]: name, kind, values (scalars), settings (every device parameter with the value used); arrays are in result.npz as <name>_<array>;
  • lossBudget, energyBudget, totals, warnings, notes, grid.

Python: res = uf.run(system, spec); res.summary, res["name_array"], res.instrument("name"); uf.load("results") opens a folder without the engine.

Branched systems

For beam splitters, interferometers, pump arms and loops, set "topology": "branched", and give every element next (Branched systems). Check the layout with ultrafast validate system.ufs --graph. Read the budget from summary["graph"]: energyClosureJ should be at rounding level.

Many runs

Use a study, not a loop: it freezes the grid (comparable results), reuses the unchanged part of the system and runs in parallel (Parameter studies). Parameter paths: element:ID/key, material:NAME/n2, pulse/energyJ, pulse/chirp/gddFs2; outputs: pulse:end/fwhmS, record:ID/BIntegralRad, instrument:NAME/key. Recorded data is core: an instrument's "measured" (compared on the recording's axis, range and resolution; result.comparisons) and measured input spectra (pulse.spectrum, with background), see Measured data. Fitting against measured data is the fitting plugin's "fit": mark transform numbers with {"value", "fit": true} or list paths; results in fit.json (layers, influence, provenance) and fitted-system.ufs. "backward" propagates a measured field to the input (Fitting).

Facts that are easy to get wrong

  • steadyState loops need gain. A passive cavity is a ringDown loop, with tolerance near the square of the accuracy you want.

  • beam.m2 is along x on a cartesian grid; beamProfile also gives m2Y.

  • Port c of a 50:50 Mach-Zehnder with equal arms is dark: all the light leaves by d.

  • SI units everywhere, with the unit in the key (lengthM, durationFwhmS, gddFs2 is the explicit exception in fs²). Wavelengths in metres: 1.03e-6, not 1030.

  • beam on an element or inputBeam in the run is needed for any nonlinearity (intensity = power / area). With a Gaussian inputBeam the beam is tracked through lenses and curved mirrors; elements with their own beam use it.

  • 1-D runs warn above the critical power for self-focusing; that is a warning about the model, not an error. Use a transverse model when spatial effects matter.

  • Gases: n2 of N2 and air is the electronic part; the rotational Raman response is separate.

  • SHG-FROG retrieval cannot determine the sign of the chirp.

  • The grid planner sizes the time window for stretched pulses and satellites (ghosts, pedestals). Override grid only with a reason; check edgeEnergyFraction in output (should be < 1e-6).

  • Synthetic data with known answers for testing an analysis: Assets.

Adaptive optics and digital twins

Two plugins, driven from Python (load them with uf.plugins.load_python_plugins(); the examples show the pattern):

  • ao: from ultrafast_ao import DmBench, Slm, DeformableMirror, SpectralShaper, Aopdf, tiled_array. A deformable mirror and an SLM are elements that compile into the core modulator; big maps are passed as a callback ("phaseTargetRad": "@name", fields={"name": f(x, y, z)}) or an NPZ file, never as JSON. The algorithms are NumPy: ultrafast.algorithms.holography (Gerchberg-Saxton, weighted GS), .control (modal and zonal reconstruction, interaction and control matrices, integrator loop) and .sensorless (spgd, modal_search, transmission_matrix). Rules that save time: choose the lens f = N dx^2 / lambda so that focal pixels are FFT bins; keep patterns in the inner half of the window; set grid.windowS to cover a shaper's side pulses (2 pi / pixel bandwidth); raise transverse.spectralSamples until spectralSamplingError is small for a pulse. Examples: AdaptiveOptics; reference page: ao.
  • twin: ultrafast twin init|fit|validate|predict|design|track|check (set ULTRAFAST_PYTHON) or ultrafast_twin from Python. A twin is the fit across conditions (the fitting plugin) plus a validation report: always mark at least one condition heldOut, read the verdict (ok, marginal, fits-training-only, unvalidated) before trusting a prediction, and give the drift tracker the expected drift in physical units. The MCP server for agents is not built. Examples: DigitalTwin; reference page: twin.

Changing the code

  • Layout: src/UltrafastEngine.Core (engine), src/UltrafastEngine.Cli, src/UltrafastEngine.Native (C API), python/src/ultrafast, plugins/ (fitting, zemax, gpu, hitran), tests/UltrafastEngine.Tests (custom runner, no NuGet).
  • Zero NuGet dependencies; the core must build for net10.0 and net48 (no newer APIs without #if NET).
  • Every element/option lives in src/UltrafastEngine.Core/Schema/BuiltIn*.cs; after changing one, regenerate ultrafast schema system -o schemas/ufs.schema.json, ultrafast schema run -o schemas/run.schema.json, ultrafast reference -o docs/reference/options.md, python python/tools/gen_types.py and python python/tools/gen_api_docs.py. Tests fail if any of these is stale.
  • Tests: dotnet run --project tests/UltrafastEngine.Tests -f net10.0 -c Release (-- --slow for all, -- Name to filter); Python: python python/tools/build_engine.py && pytest python/tests. First-time setup and which build to pick: Install and build (the "For AI agents" box). See Testing.
  • Extension points (index models, responses, elements, instruments, retrievals, compute back-ends, FFT providers) and plugins: Extending, Plugins.