Recorded by commits ac58efd through 1c2739a
Experiment Note
This was an intensive cavity-debugging day. The work included Reynolds-aware solver defaults, near-wall learning fixes, reliable checkpoint restoration, low-Re balance, suppression of a correction-bubble mode, stronger flow- orientation guards, and rebalancing toward raw physics.
A canonical soft-boundary velocity-pressure (UVP) baseline was then added. Continuity and boundary enforcement were tightened; wall objectives were targeted; and the boundary objective was made Pareto-safe.
Technical Essay
The sequence reveals the main cavity failure modes:
- Checkpoint regression: the final state could be worse than the best observed state unless best-checkpoint restoration was verified.
- Bubble mode: a local correction near the lid could create an artificial recirculation structure while improving a narrow objective.
- Orientation failure: a scalar loss could be small while flow direction was qualitatively wrong.
- Objective masking: normalized or adaptively weighted training losses could hide deterioration in raw PDE, continuity, or boundary terms.
- Formulation confounding: hard-boundary and streamfunction models could make a controller look stronger unless a canonical soft-BC UVP baseline was retained.
The response was multi-layered: use raw reference-free gates, restore the best reliable checkpoint, cap local corrections, include orientation and topology checks, and construct a baseline whose formulation is familiar and reproducible.
Reading Note
This phase reinforced the multi-objective interpretation of PINN training. A boundary term, continuity residual, momentum residual, sparse-data fit, and topology indicator need not improve together. Pareto safety is therefore more appropriate than optimizing a single weighted sum and calling it universally better.
Progress Update
The cavity solver became less fragile and the experimental language became more honest. A run would no longer count as successful merely because it converged. It had to pass raw-physics, topology, speed, boundary, and checkpoint gates.