Skip to content

Workflow

This page describes the standard PME-toolkit workflow from input data to outputs.

PME-toolkit is organized as a configuration-driven pipeline:

  1. define the input dataset
  2. define the workflow in a JSON configuration file
  3. run dimensionality reduction
  4. inspect results and visualizations
  5. optionally perform backmapping

1. Input dataset

A PME workflow starts from a dataset containing one or more of the following components:

  • geometry
  • design variables
  • physical information

These components must be sample-aligned, i.e. each sample must refer to the same design configuration across all data sources.

The dataset is typically stored in an external .mat file and referenced in the configuration through:

"io": {
  "dbfile": "path/to/database.mat"
}

See the Datasets and Input data pages for details.

2. JSON configuration

The workflow is fully defined through a JSON file.

A standard run configuration specifies:

  • method (mode)
  • retained variance (CI)
  • geometry definition (geom)
  • design variables (vars)
  • physical quantities (phys, if applicable)
  • filters (filters)
  • input/output settings (io)

This JSON file is the central entry point of the workflow.

See the Configuration specification page.

3. Preprocessing

Before dimensionality reduction, the dataset may be filtered and normalized.

Available preprocessing steps include:

  • NaN removal
  • goal-oriented filtering
  • IQR-based filtering
  • variable-range handling

These operations are controlled in the filters and vars sections of the configuration file.

4. Dimensionality reduction

PME-toolkit supports three main modes:

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

The workflow constructs the embedding from the data described in the configuration and retains the number of modes required to satisfy the confidence index:

CI = retained variance threshold

Typical values are:

  • 0.95
  • 0.99

The result is a reduced latent representation of the original design space.

5. Outputs

A standard run produces:

  • reduced coordinates
  • retained dimensionality
  • eigenvalues and modes
  • reconstruction metrics
  • saved model
  • visualization outputs

Results are typically written under:

results/

depending on the io.outdir setting.

6. Visualization

Visualization is part of the standard workflow and helps interpret the quality of the embedding.

Typical outputs include:

  • scree plot
  • retained variance
  • variance by source
  • NMSE by source
  • variable participation
  • geometric modes

See the Visualization page.

7. Backmapping

After a model has been generated, reduced coordinates can be mapped back to the original variable space through a separate backmapping workflow.

Backmapping requires:

  • the original case configuration
  • reduced coordinates input
  • a dedicated backmapping JSON file

Run it with:

Python

pme-back tests/cases/test_glider_back.json

MATLAB

run_back("tests/cases/test_glider_back.json")

See the Backmapping and Backmapping configuration pages.

8. Typical execution

Python

pme-run tests/cases/test_glider.json

MATLAB

run_pme("tests/cases/test_glider.json")

These commands execute the full workflow on the self-contained glider reference case.

Summary

The PME-toolkit workflow is:

  • data-driven
  • configuration-driven
  • reproducible
  • modular

The same logic applies to standard tests, benchmark cases, and externally hosted datasets.