Reference simulation

Learning goals¶

• Build a transparent numerical SMIB reference with an event schedule

• Document solver settings, failures, and scenario-based data splits

• Treat the simulator as scientific ground truth for later ML comparisons

Reading¶

• Lab 1 — reference simulation

• Numerical ODE notes in Preparation

Why a reference first¶

Before PINNs or surrogates, you need:

1. Equations and parameters in one place

2. An event log (fault inception, clearing, horizon)

3. A convergence / failure policy (do not silently delete failed runs)

4. A manifest: scenario IDs, seeds, solver tolerances, units

Machine learning cannot repair an undocumented dataset.

Minimum deliverables (Lab 1)¶

• SMIB integrator (teaching code in src/sciml_ps)

• At least two numerical settings compared (e.g. step size or method)

• Dataset split by scenario, not by randomly shuffling time points across events

• Flagged numerical failures retained as records

Worked example¶

Compare trajectories for the same physical scenario at two time steps. Report max $|\delta|$ difference and whether the stability label (if defined) flips.

Lab¶

• Lab 1 — Reference simulation and scientific dataset

Further reading¶

• Lab 1 protocol and dataset record fields

Common misconceptions¶

• Deleting “ugly” trajectories that failed to converge

• Splitting every time sample randomly into train/test (leaks the same event)

• Calling ANDES optional until network labs — classical SMIB does not require ANDES