Repository navigation
Examples - BUGFIX - Re-tune the Solovev wall fixture and close out follow-ups from the surface-sampling review #544
Description
Follow-ups from @d-burg's review of #398 (sample every flux surface at the same straight-fieldline (SFL) angles). He marked these as fine to leave for after the merge.
1. The Solovev fixture is now a ~1000:1 cancellation
After #398, Solovev_ideal_example gives et[1] = 0.0109 against ep = −10.42 and ev = +10.43. Any later change that moves ep by 1e-4 relative moves et[1] by ~10 %, so solovev_n1 and the Solovev kinetic cases will amplify every future change.
- Re-tune
[Wall] standoff_fraction(currently 0.2415) so et[1] sits clearly away from marginal, then re-pin the Solovev harness baselines. - Fix the deck comment: it still says "with this wall, et[1] = +0.24 (barely stable)". That value came from the resampling noise Equilibrium - BUGFIX! - Sample flux surfaces on common straight-fieldline angles, cap core EL knots, and suppress near-axis kinetic terms #398 removed.
2. Is the ~2500× NTV torque move in solovev_kinetic_ntv resonant amplification?
Hypothesis from the review: sitting this close to marginal amplifies the plasma response by roughly (0.677 / 0.0107)² ≈ 4000. Comparing the response amplitude between develop and #398 at the same applied field would settle it. Re-tuning the wall (item 1) should shrink the effect.
3. Put the dense-output samples exactly on ψ = ψ₀
direct_fieldline_int's refinement callback saves pre-refinement states (save_positions=(true, false)). The dense interpolant is therefore offset by one step's refinement, and the SFL samples sit slightly off the surface, at tolerance level. Calling direct_refine on each sample would put them on the surface exactly. Not blocking: measured on Solovev, traced points lie within 1.4e-5 in ψ of the exact surfaces, which is the floor set by the ψ(R, Z) spline.
4. PENTRC-experienced review of #414's physics
#414 (near-axis kinetic validity suppression) is merged into #398. @ebursch asked for a look from someone with PENTRC experience, which is still open.
🤖 Generated with Claude Code
Activity
logan-nc commentedon Oct 24, 2026
Re item 3 (put the SFL samples exactly on ψ=ψ₀): PrincetonUniversity/GPEC#355 proposed the same thing for the Fortran code —
direct_refineon each common-angle sample before evaluating r² and the Jacobian, leaving the integrator state and θ(η), φ untouched. We closed it there (not a bug fix; O(etol) effect; v1.6 release day) and are treating this issue as the home for it.Two points from that review worth carrying over: (a) the case for doing it is high mpsi — per-surface radial drift is uncorrelated between surfaces, so ∂(r²)/∂ψ from the spline amplifies it by ~1/Δψ and the metric derivatives get noisier as the radial grid is refined, not smoother; (b) the fork measurement behind #355 (krystophny/GPEC#2) claims a ~20× drop in reconstructed Bpol error at mpsi=512, etol=1e-8 on a TC24 g-file, which is more than the mechanism predicts and would be worth reproducing with the round-trip ψ residual as the metric. Note the current callback feeds the refined radius back into the integrator (
refine_affect!), which Fortran never did, so the Julia and Fortran charts already differ at tolerance level here.