Skip to content

Configuration specification

PME-toolkit uses JSON configuration files to define dimensionality reduction workflows.

This page documents the actual configuration schema used by the toolkit.

Full example

{
  "mode": "pi",
  "CI": 0.95,
  "baseline_col": 0,
  "geom": { ... },
  "vars": { ... },
  "phys": { ... },
  "filters": { ... },
  "io": { ... }
}

Top-level fields

mode

Type: string

Defines the method:

  • "pme" → geometry + variables
  • "pi" → physics-informed PME
  • "pd" → physics-driven PME

CI

Type: float

Confidence index (retained variance), e.g.:

0.95 → 95% retained variance

baseline_col

Type: integer

Index of the baseline configuration in the dataset.

Used for:

  • normalization
  • fixed-variable handling

Geometry (geom)

Defines the geometric structure.

Fields

  • Jdir (int)
    Number of spatial directions per point
    (e.g. 2 → 2D, 3 → 3D)

  • patches (list)

Each patch:

{
  "name": "P1",
  "K": 784
}

Where:

  • name → identifier
  • K → number of points in the patch

Notes

Total geometric dimension:

n_geom = Jdir × sum(K over patches)

Variables (vars)

Defines parametric design variables.

Fields

  • Mbase (int)
    Total number of variables

  • idx_active (list)
    Indices of active variables

    [] → all variables active

  • fixed_policy (string)

    "baseline" → fixed variables set to baseline values

  • Urange_file (string)
    Path to variable bounds

  • use_db_range_if_missing (bool)
    Fallback to database if range file is missing

Physics (phys)

Defines physical data used in PI-PME / PD-PME.

Fields

fields

List of distributed fields:

{
  "name": "Cp",
  "nCond": 1,
  "disc": { "K": 784 }
}

Where:

  • name → field name
  • nCond → number of conditions
  • disc.K → discretization size

scalars

List of scalar quantities:

{
  "name": "D",
  "nCond": 1
}

Notes

  • fields → spatially distributed data
  • scalars → global quantities

Filters (filters)

Defines dataset filtering.

Fields

remove_nan (bool)

Remove samples containing NaN values.

goal

{
  "enable": true,
  "metrics": [
    {
      "c_offset": 0,
      "rule": "positive"
    }
  ]
}
  • enable → activate filter
  • metrics → list of constraints

Each metric:

  • c_offset → column index in scalar outputs
  • rule → filtering rule (e.g. "positive")

iqr

{
  "enable": true
}

Apply interquartile range filtering.

Input / Output (io)

Defines data access and output behavior.

Fields

  • dbfile (string)
    Path to database file (.mat)

  • transpose_if (string)

    "never" | "always" | "auto"

Controls data orientation.

  • outdir (string)
    Output directory

  • save_model (bool)
    Save trained model

Backmapping configuration

Backmapping uses a different JSON structure:

{
  "model_path": "results/model.mat",
  "x": [ ... ]
}

Where:

  • model_path → trained model
  • x → reduced coordinates vector

Key constraints

  • all data must be sample-aligned
  • geometry, variables, and physics must share sample count
  • x dimension must match retained modes
  • paths must be valid and accessible

Minimal working example

{
  "mode": "pme",
  "CI": 0.99,
  "baseline_col": 0,
  "geom": {
    "Jdir": 3,
    "patches": [
      { "name": "P1", "K": 100 }
    ]
  },
  "vars": {
    "Mbase": 10,
    "idx_active": [],
    "fixed_policy": "baseline"
  },
  "filters": {
    "remove_nan": true
  },
  "io": {
    "dbfile": "database.mat",
    "outdir": "results",
    "save_model": true
  }
}

Summary

The configuration file fully defines:

  • data structure
  • preprocessing
  • method
  • output

and is the central interface of PME-toolkit.

Backmapping uses a separate configuration format (see Backmapping page).