Repository navigation
Automatic cold Picard warm-up diverges a LINEAR constrained Stokes solve (Stokes_Constrained + multiplier free-slip) #817
Description
On `bugfix/picard-warmup-791` with the consistent tangent as the default: the constrained free-slip annulus of `tests/test_1061_constrained_freeslip.py::contrast_annulus` (ViscousFlowModel, smooth lateral viscosity ramp 1..1e3, pressure nullspace, basic line search, no `snes_type` set) gives
| setting | result |
|---|---|
| `consistent_jacobian=False` (Picard throughout) | CONVERGED_FNORM_RELATIVE, 1 it, 3 KSP, 1.3e-9 |
| `consistent_jacobian=True`, `picard=-1` (no warm-up) | CONVERGED_FNORM_RELATIVE, 1 it, 3 KSP, 1.3e-9 |
| `consistent_jacobian=True`, automatic warm-up (one Picard step, then Newton) | DIVERGED_DTOL, 1 it, 9 KSP, 6.7e-2 |
The problem is linear, so the frozen and consistent tangents are the same operator; the warm-up path (blended kernel compiled with the alpha atom, one `snes.solve` at rtol 0 / max_it 1, tolerances restored, second solve) is what differs, and on the constrained saddle point (velocity | [p, h] grouped Schur, grad-div augmentation 1e2) it diverges where either plain path converges in one iteration. Probe: `scratchpad/constrained_probe.py` in the campaign session.
Mitigation on the branch: the automatic step is now skipped when the flux coefficient does not depend on the unknowns (`_flux_is_linear_in_unknowns`) and for `snes_type=ksponly`. A NONLINEAR constrained problem would still take the warm-up through this path, so the defect itself stands.
Activity
Nonlinear solver mechanisms: Newton warm-up, residual-keyed blend, pa…
lmoresi commentedon Oct 6, 2026
The review of PR #794 found this issue misdiagnosed. The automatic Picard warm-up is not the cause:
Stokes_Constrainedfails on its second Newton iteration whatever the tangent.We ran the test_1061 contrast annulus with η·(1 + 1e-6 ε̇_II), which the linearity guard classes as nonlinear:
DIVERGED_DTOLafter 1 iterationpicard=-1(no warm-up)DIVERGED_MAX_ITat 50; step 1 reaches 8.5e-7, then it oscillates between 3.9e-5 and 6.6e-2consistent_jacobian=False(the old default)PR #794 keeps the guard that skips the warm-up for a flux whose coefficient does not depend on the unknowns. That covers the linear case this issue was filed on. It does not fix the constrained solver's nonlinear iteration, which is the actual defect.
Suggest retitling to "Stokes_Constrained diverges on the second Newton iteration of any nonlinear flux (multiplier free-slip)".
Underworld development team with AI support from Claude Code