Recorded by commits 0e24545, d7639f1, 99a8f77, 9176489, and bdc97fb
Experiment Note
A cavity patch-aware local controller was added and merged. This focused the adaptation on spatially localized defects such as the lid band, corners, and interior recirculation regions rather than treating the square cavity as one uniform domain.
An Adam-to-L-BFGS optimizer schedule was added, then reverted within minutes. The short lifetime of this change is informative: L-BFGS can improve final PINN fits, but line searches use a variable number of closure evaluations. That makes exact fixed-step or equal-collocation accounting difficult.
Technical Essay
The cavity is geometrically simple but numerically hostile. The top wall moves, the other walls do not, and the idealized top corners contain discontinuous boundary data. A regional controller therefore needs to distinguish genuine interior flow defects from corner singularity behavior. Patch-aware scores made that separation possible, but also increased the risk of overreacting to a small pathological zone.
The optimizer revert established a fairness principle used later: if the claim is about a controller, compared methods must share the same optimizer schedule and effective budget. A final optimizer polish can be studied separately, but it should not be allowed to obscure where the improvement originated.
Reading Note
The cavity benchmark literature emphasizes centerline profiles, vortex centers, and flow topology rather than one aggregate error alone. Optimization studies in PINNs also show that Adam and L-BFGS play different roles. The resulting note was to report optimizer-stage effects explicitly and disable final repair in strict equal-compute studies.
Progress Update
The local controller became more geometry-aware. The project also learned to treat optimizer choice as part of compute accounting rather than an innocent implementation detail.