Recorded by commits d91f1a1 through 7e93b72
Experiment Note
A method-isolated Reynolds-continuation workflow was created for the cavity. Full-field cavity comparison artifacts were added from external CFD sources, with post-hoc evaluation kept separate from training. Publication-oriented fairness studies were added: same schedule, same model steps, same collocation evaluations, and same wall-clock budget. Modern baselines included vanilla, RAR, self-adaptive attention, gradient balancing, and gradient-enhanced PINNs.
Technical Essay
Continuation addressed a practical cavity bottleneck: jumping directly to a high Reynolds number can destabilize training or produce an unphysical flow topology. A method-specific chain can warm-start Re=150 from Re=100, then move through increasing Reynolds numbers. To keep the comparison legitimate, each method must have its own chain, optimizer moments must be reset at a new stage, and any previous-Re replay must be fixed, seeded, reference-free, and identical across methods.
Full-field CFD data solved an evaluation limitation. Ghia profiles can verify selected centerlines, but they cannot provide whole-field L2 error or spatial error maps. The dense CFD field was therefore valuable post hoc. It remained forbidden as a hidden controller signal except in explicitly labeled sparse-CFD reconstruction experiments.
Fairness required more than equal epochs. VARA performs diagnostics, candidate ranking, probes, and sometimes rollback. Equal model updates, equal collocation evaluations, and equal wall time answer different questions. The repository began logging these budgets separately instead of claiming one universal notion of equal compute.
Reading Note
The baseline survey expanded to self-adaptive pointwise weights, gradient-based global balancing, gradient-enhanced residual constraints, residual attention, ReLoBRaLo, and causal approaches. Each baseline represented a distinct kind of adaptivity. The important comparison rule was to preserve its independent mechanism rather than implement all baselines as superficial configuration variants of VARA.
Progress Update
The research moved toward publication-grade protocol design. The experiment suite could now separate controller gains, continuation gains, reference-based evaluation, and compute-budget effects.