Skip to content

studies.md

Latest commit

 

History

History
74 lines (60 loc) · 3.85 KB

studies.md

File metadata and controls

74 lines (60 loc) · 3.85 KB
Raw

Parameter studies

A study runs a system many times with parameters varied: scans, optimisation and sensitivity analysis.

// study.json
{
  "format": "ultrafast-study",
  "system": "line.ufs", "run": "run.json",
  "mode": "scan",                                    // scan | optimize | sensitivity
  "sampling": "sobol", "samples": 128, "seed": 1,    // scan: grid | random | lhs | sobol
  "parallel": 4,                                     // points run at the same time
  "parameters": [
    { "path": "element:plate/lengthM", "min": 2e-3, "max": 8e-3 },
    { "path": "pulse/energyJ", "min": 1e-6, "max": 1e-4, "scale": "log" },
    { "path": "material:FusedSilica/n2", "min": 2.2e-20, "max": 3.0e-20 }
  ],
  "outputs": ["pulse:end/fwhmS", "record:plate/BIntegralRad", "instrument:xc/prePulseContrastDb"]
}
ultrafast study study.json -o study-results         # study.json + study.csv (one row per point)
r = uf.study(system, spec, parameters=[{"path": "element:plate/lengthM", "min": 2e-3, "max": 8e-3, "points": 7}],
             outputs=["record:plate/BIntegralRad"], mode="scan")

Parameters

path changes
element:ID/key an element value, through an override (anything an override can change: lengths, angles, n2, coatings by name ...)
element:ID/a/b a nested value of the element (edited in the system JSON), e.g. element:cell/beam/w0M
material:NAME/property a material property (override): n2, ...
pulse/key[/sub] the run's pulse, e.g. pulse/chirp/gddFs2
run/a/b, system/a/b anything else in the run spec / system file

Each has min and max (scale: linear or log), or a list of values; points for grids; nominal (default the middle) for the reference run and local sensitivities.

Outputs

output value
pulse:LOC/metric pulse summary at an element id, input or end: fwhmS, peakPowerW, energyJ, spectralFwhmM, ...
record:ID/field element record: BIntegralRad, EnergyOutJ, ArrivalS, PeakFluenceJPerM2, a loss mechanism, or a value key (trackedBeamRadiusXM, peakElectronDensityM3, ...); out.metric for the summary after the element
instrument:NAME/key any scalar an instrument reports
energy:LOC energy at a location

Modes

  • scan: grid (every combination of points per parameter), random, lhs (Latin hypercube), sobol (low-discrepancy, up to 16 parameters).
  • optimize: "objective": {"output": "...", "goal": "min" | "max" | "target", "target": v, "maxEvaluations": 200} -- Nelder-Mead on the scaled box. The result has analysis.optimum.
  • sensitivity: "sensitivity": {"method": "local", "stepFraction": 0.01} (central differences: derivative and elasticity d ln y / d ln p at the nominal point) or {"method": "sobol", "samples": 256} (global first-order and total indices, Saltelli / Jansen estimators, N (d + 2) runs).

Speed

  • The grid is frozen at the nominal run's plan (freezeGrid, default true), so every point uses the same samples and results are comparable.
  • Whole arrays (arrayOutputs): ["spec/spectrum"] keeps an instrument's array per point in study.npz, with one row per point. Give output to write it. Example 34 builds a FROG trace this way.
  • Prefix caching (prefixCache, default true): when every varied parameter acts on elements after some index, each point restarts there from the nominal run's pulse instead of recomputing the elements before.
  • parallel runs points at the same time (each run also uses multiple cores for its FFTs).

Tests: SpeedStudyTests (B ∝ L in a scan with and without the cache; optimisation of a target B; elasticities of B to L and n2 equal 1; Sobol indices of an influential and a non-influential parameter; stratification of the Sobol sequence).