Physics Validation Lab (day 24)¶

Is this model or result physically and numerically consistent? Live at https://validate.pinneapple.org. Source: apps/validation_lab.

Report of the fixtures with one defect each: three groups, per-target status

Upload predictions or simulation results. Describe the equations, parameters, units, bounds, boundary conditions and reference data in a form, or upload a pinneapple check configuration, or reference a zoo model with query points. The lab runs the physics checks and reports them in three groups that answer different questions:

Group	Question
Numerical verification	Were the equations solved right? Residuals, finite values, intact artefacts, method status
Validation against data	Does it match trusted data, inside the range the data covers? Reference error, QoIs, out-of-range queries
Physical consistency	Conservation, boundary and initial conditions, bounds and constraints, units

For each target, the report says which groups did not run (for example, no reference means no validation against data). Each check that ran has an evidence record: the SHA-256 of its inputs, its outputs, and start and finish times (evidence ledger). The report also holds the configuration and the hash of every input file, so pinneapple check --config pinneapple-check.toml reproduces it.

No single test guarantees that a model represents reality, and neither does this set of tests. A pass means only that the listed checks ran and held on these inputs, with these tolerances.

A zoo model queried outside its training range: the validation group fails

On the mesh (3-D results)¶

When a result is 3-D and lives on a mesh (an OpenFOAM case zip, a VTU/VTK file with cells) and has a reference, the report adds a 3-D view of what the reference check (PC701) compared: the prediction, the reference and |prediction − reference| on the mesh, each with a colour bar. A displacement field deforms the body (scale factor, true scale). Declared boundary conditions are marked (OpenFOAM patches, planes such as x=min). A list gives the regions with the largest error: connected points in the top 10 % of the error, with their maximum and mean, size, share of the squared error and the faces of the domain they touch. Click a row to zoom to it. The relative L2 shown is the PC701 number. 2-D results, tables and models keep the current views. Library: pinneapple_veriphysics.mesh_views, pinneapple_analysis.validation.error_map, pp.viz.field_view.

The 3-D example is chosen to fail: the ROM Lab's POD + GPR model (4.4 % mean held-out error) at its worst held-out design, near the edge of its training box. It shows a failing reference check, not the ROM's typical quality.

The ROM Lab's model at its worst held-out design against the FEM: |prediction − reference| of the von Mises stress on the deformed cantilever, with the regions of largest error at the clamp

The displacement error of the same model: one region, at the loaded tip

Automate it¶

POST /api/validate takes files plus config (TOML) or spec (JSON) and returns the report. The interactive docs are at /docs.
In CI, the GitHub Action PINNeAPPle-Labs/PINNeAPPle@main runs the same checks on every pull request.
The report JSON, with its ledger, feeds the Engineering Evidence Report (day 30).

The form: a temperature field checked against the heat equation, its boundary value, units and bounds

Limits¶

Python models and checkpoints are not run on the server. Upload their predictions, or use the Action.
The default tolerances of pinneapple check decide pass or fail unless you set them per target.
The report is pinneapple-validation-lab/v1. Export to validation-report/v1 (pp.validate) and its credibility levels will follow once those PRs are merged.

Monetization: validations per model, dataset or version, with private projects and CI history for teams.