Recorded by commits 3653e16 and 38b0d49
Experiment Note
Near-wall cavity repair was integrated with VARA V2. Re=100 speed and topology stabilizers were tuned. The goal was not only to reduce residuals but to prevent unphysical flow speed, bad primary-vortex orientation, and near-lid failures.
Technical Essay
The cavity bottleneck concentrated near the moving lid and corners. Uniform training under-resolved this region; aggressive local repair could overfit the lid band and create an interior correction bubble. The solution was a guarded near-wall phase with bounded action size and explicit topology, speed, continuity, and boundary checks.
A key lesson was that "repair" cannot mean minimizing the same weighted loss used during training. Adaptive weights can make the weighted objective look better while raw physics worsens. The reliable gates therefore emphasized unweighted, reference-free diagnostics.
Reading Note
Classical cavity solutions make the physical flow structure visually intuitive: one expects a correctly oriented dominant vortex with Reynolds-dependent secondary structures. This supported using flow orientation and vortex sanity as qualitative-quantitative gates alongside residual norms.
Progress Update
The low-Re base stage became more trustworthy. Reliable continuation was being treated as a chain of validated physical states rather than a sequence of checkpoints that merely finished training.