Recorded by commits 63598d1 through 8ed8755
Experiment Note
The isolated Taylor-Green VARA V2 wiring was repaired and evaluation was optimized. A matched vanilla comparison and stress-sweep configs were added for Re=100, 500, and 1000 at multiple step, sample, and batch budgets.
An isolated PDE-generalization suite was then added with manufactured 2D Burgers, Allen-Cahn, and advection-diffusion benchmarks. These trainers, diagnostics, residuals, plots, and summaries were separated from the cavity implementation so the same controller logic could be tested without relying on Navier-Stokes-specific plumbing.
Technical Essay
Taylor-Green exposed the difficulty of extending a steady-flow controller to a transient vortex decay problem. Space-time sampling, initial conditions, and time-dependent residuals must be handled consistently. "Isolated wiring" was important: legacy cavity assumptions, metric names, or reference access could not be allowed to enter the transient controller.
The manufactured PDE suite was a stronger generalization test. Exact fields and forcing terms were defined analytically, but the controller used only permitted training signals. Patches extended through space-time, enabling the controller to follow a moving Allen-Cahn interface or transported advection-diffusion layer.
Reading Note
Allen-Cahn was selected as an interface-dominated reaction-diffusion equation; advection-diffusion represented a transported localized layer; Burgers provided nonlinear transport. Causal PINN literature was relevant because transient errors can propagate from early time, although VARA's mechanism remained adaptive space-time patch selection rather than a fixed causal schedule.
Progress Update
The controller was no longer only a Navier-Stokes method. The software and experimental design now supported a cross-PDE claim, with strict isolation tests for residuals, reproducibility, and controller inputs.