Recorded by commits ab238d0 through 1994831
Experiment Note
The lid-band momentum coupling was fixed. A lifted UVP formulation was added for Re=100, followed by an attempted divergence-compatible velocity lift that was reverted. Sparse CFD supervision was introduced for cavity reconstruction, using a restricted observation pool rather than the dense evaluation field.
An opt-in sparse-polish phase was developed. Continuity was tightened, VARA resampling core counts were fixed, and non-regression/Pareto safety was added. A gate-rescue mechanism was implemented then reverted. A curriculum-controlled sparse polish was retained. CFD-reference resolution for continuation was tried and reverted. Sparse-loss wiring for adaptive baselines was fixed, reverted, and then rebuilt safely.
Technical Essay
Sparse reconstruction was one of the hardest integration points because it combined four competing requirements:
- Improve the measured velocity field from only a small observation fraction.
- Preserve continuity, momentum, boundary conditions, and flow topology.
- Give every baseline access to exactly the same sparse observations.
- Prevent the dense CFD evaluation field from leaking into control decisions.
The repeated reverts show why loss wiring mattered. An adaptive baseline might construct its loss dictionary differently from vanilla or VARA. Adding sparse polish at the wrong layer could omit the term, apply it twice, change only one method's gradient path, or count a different number of collocation evaluations. The safe construction centralized the sparse term, kept it opt-in, and tested disabled behavior for legacy equivalence.
The gate-rescue revert is another useful lesson. A mechanism that tries to save a nearly rejected action can quietly weaken the meaning of the guard. The final preference was conservative: curriculum and bounded Pareto tolerances, not a special exception that turns failure into acceptance.
Reading Note
Sparse flow reconstruction work, including hidden-fluid-mechanics and sparse PIV/PTV assimilation, clarified the application setting. Physics provides a powerful prior when observations are limited, but identifiability and residual tradeoffs remain. The paper therefore framed VARA as reconstruction-first in the cavity rather than claiming exact incompressibility from a soft UVP model.
Progress Update
Sparse CFD reconstruction became a first-class experiment. The debugging cycle also hardened the scientific software: all methods shared the same data pool, adaptive baselines received correct losses, and rejected rescue shortcuts were removed.