Repository navigation
The :curved reference wing does not converge: LOOP returns NaN gamma at 5.7° and 10° #395
Description
The :curved reference wing (the one test/thesis_oriol_cayon.jl builds) does not converge: solve! with the VSM LOOP solver returns FAILURE, with NaN in lr.gamma_new, at α = 5.71° and at α = 10°, |va| = 20. Measured on main at 2b312ea (2026-09-16) while working on #317, with N = 40 panels. Its test checks geometry only and never solves, so the suite does not see it.
ba = ref_wing(:curved) # the thesis_oriol_cayon.jl curved wing, N = 40
set_va!(ba, 20.0 .* [cosd(5.7106), 0.0, sind(5.7106)])
s = Solver(ba; aerodynamic_model_type=VSM)
sol = solve!(s, ba)
sol.solver_status, s.lr.converged, count(isnan, s.lr.gamma_new) # FAILURE, false, > 0What I would do: first find whether the NaN comes from the geometry (a panel frame or AIC entry) or from the iteration, then add a solve to the :curved test so the reference wing is pinned to converge. Not reproduced on current main yet; this records it so it is not lost. #315 is a different LOOP convergence failure (the fixed point at α = 0).
Activity
1-Bart-1 commentedon Sep 30, 2026
Check if it does converge in the python version
1-Bort-1 commentedon Sep 30, 2026
Fixing all of it in this PR, on top of 8c8db26 (halving + stop at non-finite): finding why solver_test_wing refined to 12+ panels does not converge (LOOP and NONLIN) and fixing that in the solver, then finding why the unrefined cl/cd/cm aggregation misses the panel sums by 17 % and fixing whichever side is wrong. The unrefined-consistency test then runs its checks for real.
1-Bort-1 commentedon Sep 30, 2026
Question from
VortexStepMethod.jl-395Which fix do you want for #395? The NaN is not the geometry. The AIC, panel frames and initial gamma are all finite. The LOOP iteration diverges: max|gamma| is 3.8e50 by iteration 10 and NaN by 20.
Why: NONLIN finds the fixed point (VSM CL 0.4223, residual 1e-13). At that point the LOOP map's Jacobian has an eigenvalue at −210, from the 1.3 cm wide cosine-spaced tip panels on a 1 m chord. A relaxed step is stable only for ω < 2/(1+|λ|) = 0.0095. The default 0.03 and the single retry at 0.015 both diverge. LLT's eigenvalue is −106 (limit 0.019), which is why LLT converges on its retry. At ω = 0.0075 VSM converges in about 1220 iterations at both 5.7° and 10°.
Options:
A. Solver fix:
solve_base!keeps halving ω while ω > 1e-3, which its guard already hints at, and LOOP stops as soon as gamma goes non-finite instead of running 1500 NaN iterations. This fixes every wing this stiff. The cost is up to ~5 retries on cases that never converge.B. Test only: add a solve to the
:curvedgeometry test withsolver_type=NONLIN(or an explicit ω = 0.0075) and pin CL. The default LOOP still fails on this wing.C. Both.
I recommend A: the class is 'stiff wing, fixed ω' rather than this one wing, and a test that converges only on NONLIN hides the failure. I'd also pin the
:curvedsolve with the default solver, so it goes red today and green after the fix.1-Bart-1 commentedon Sep 30, 2026
A, should warn when halving, and it errors when NaN already right? SymAWE catches this error.
16 remaining items
1-Bort-1 commentedon Nov 5, 2026
From
VortexStepMethod.jl-410(#410), tell 33, 2026-11-05T00:25:24ZTask #410 makes the default relaxation factor 0.8 of the stability limit from Python v2.3.0 (nothing as default, explicit numbers still honoured) and copies your area-weighted unrefined aggregation and polar row into its branch, because its converging solves exposed the same bug. On that default your curved wing converges. Your per-panel cap remains useful for explicit factors and your non-finite stop is not in #410. Expect a conflict in solver.jl and test_unrefined_dist.jl: the aggregation hunks are identical to yours.
1-Bort-1 commentedon Nov 5, 2026
On tell 33 from
VortexStepMethod.jl-410: No change to #402, it stays at c62bd58: its cap and non-finite stop still apply to explicit factors. A trial merge with your 242880a conflicts only in src/solver.jl; test_unrefined_dist.jl merges clean. Whichever lands second resolves it.