Repository navigation
fix(torsion): coupled held lines step, body passes converge, a lazy wave in current starts at any pretwist - #82
Conversation
…ave in current starts
Three torsion defects of the 0.4.0 robustness campaign (all on dev):
- Strong torque-tension coupling: with K >= 0.85 sqrt(EA GJ) (and -0.8
with one end free) a held line in exact static equilibrium failed its
dynamic step ("||G|| = 2.4e-8 > tol = 1e-8"): the round-off floor of the
convergence norm took only the kernel's EA - K^2/GJ. It now takes the
coupling's share of the axial force and of the element torque (whose
twist is formed from coordinates). Gate torsion_coupling_still: K up
to 0.99 sqrt(EA GJ), either sense, one end free; fails on the old code.
- Static body passes of a cable restrained in torsion between two Rigid6
bodies stalled near 2e-9 at 360 deg (Rigid) and 300 deg (Pinned): the
1e-10 pass tolerance sat below the resolution of the core's body
balance. The passes now stop at that resolution (four times the core's
residual load), and the failure advice no longer suggests bodyIC deck.
Gate torsion_body_passes: the pretwist sweep converges (Rigid 360 deg
in 6 passes, torque 4886.24 N m); fails on the old code.
- A lazy wave with 25 modules restrained in torsion in a 0.5 m/s in-plane
current was refused by the torsion stage ("stalled at Phi = 0") although
the twist added no negative stiffness mode to its untwisted state,
which is unstable out of its plane and accepted by the deck without
torsion. A twisted state with no more negative modes than its untwisted
one is now kept and named as such. Gate torsion_current_start (pretwist
0 and 360 deg); fails on the old code.
Also: the "unresolved/kinked equilibrium" refusal names the factor to
refine the section's NumSegs by.
Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
…BLAS kernel The start of the lazy wave restrained in torsion in an in-plane current depended on the pretwist and on the round-off of the BLAS kernel: - The static helper builds (coarse branch selection, auxiliary-weight seed, friction reference, no-current diagnosis) ran the torsion stage, and the twisted coarse state seeded the untwisted solve on the deck mesh. Pretwist 180 deg stopped in the attachment polish; 360 deg needed the polish's fallbacks and stopped with OPENBLAS_CORETYPE=Prescott. Helper builds now solve the untwisted line. - Each pass of the attachment fixed point after the first applied the drag of the last twisted shape in full at the untwisted start, which pushed the out-of-plane-unstable untwisted state away before any twist (270 deg and more). The load change is now ramped with the imposed twist (f_nodal_twist of CD_HermiteCable_Static_Solve). Pretwists 0 to 720 deg now start without fallbacks, with the same torques on the Haswell and Prescott kernels, proportional to the pretwist. torsion_current_start sweeps 0/180/360/720 deg and checks both. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Codex Review: Didn't find any major issues. Can't wait for the next one! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
The pretwist sweep runs four independent decks; one after another they took 322 s in a local Debug build and passed the 600 s test timeout on the CI Debug runners. They now run concurrently (158 s locally in Debug, 26 s in Release), and the timeout is 900 s, as for the other long driver tests. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
|
Codex Review: Didn't find any major issues. Another round soon, please! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Summary
Three torsion defects from the 0.4.0 robustness campaign (all present on dev), each with a gate that fails on the old code.
torsion_coupling_still). WithK >= 0.85 sqrt(EA GJ)(and-0.8with one end free in torsion) a held line in exact static equilibrium failed its dynamic step within 0.05-0.9 s ("Newton did not converge, ||G|| = 2.4e-8 > tol = 1e-8"). The round-off floor of the step's convergence norm took only the kernel'sEA - K^2/GJ; it now also takes the coupling's share of the axial force and of the element torque, whose twist-(K/GJ)(l - l0)is formed from coordinates. Held lines withKup to0.99 sqrt(EA GJ), in either sense and with one end free, hold at the defaultdtMand tolerance.torsion_body_passes). The passes compared the change of the cable's end loads with 1e-10 of their scale, below the resolution of the core's body balance (1e-9 of the bodies' net weight): at 360 deg (Rigid ends) and 300 deg (Pinned) they stalled near 2e-9 and the deck stopped, with advice (bodyIC deck) that was not the remedy. They now stop at the resolution the core leaves; the Rigid 360 deg case converges in 6 passes (torque 4886.24 N m) and the pretwist sweep 120-720 deg converges for both end kinds.torsion_current_start). With 25 buoyancy modules in a 0.5 m/s current its untwisted static state, solved in its vertical plane, is unstable out of that plane on its mesh, which the deck without torsion accepts with a note; the torsion stage refused the twisted state although the twist added no negative stiffness mode ("torsion continuation stalled at Phi = 0 (target 0)" at pretwist 0, an unstable equilibrium after a failed Newton solve at 360 deg). A twisted state with no more negative modes than its untwisted one is now kept, and the note names the instability as the untwisted state's. Two more causes made this start depend on the pretwist and on the BLAS kernel's round-off. First, the coarse branch-selection build (and the other static helper builds) ran the torsion stage, and its twisted state seeded the untwisted solve on the deck mesh: 180 deg stopped in the attachment polish, and 360 deg needed the polish's fallbacks and stopped withOPENBLAS_CORETYPE=Prescott. Second, each pass of the attachment fixed point after the first applied the drag of the last twisted shape in full at the untwisted start, so 270 deg and more stopped before any twist. Helper builds now solve the untwisted line, and the change of the attachment loads is ramped with the imposed twist (newf_nodal_twistofCD_HermiteCable_Static_Solve). Pretwists from 0 to 720 deg now start without fallbacks, with the same torques on Haswell and Prescott kernels, proportional to the pretwist.Also: the "unresolved/kinked equilibrium" refusal names the factor to refine the section's
NumSegsby.Checks
ctest 356 of 356 (OpenBLAS
Prescottkernel); checked Debug build (-fcheck=all) on the 26 torsion tests;torsion_current_startdeck at pretwist 0-720 deg on the Haswell and Prescott kernels; example A/B 53 of 53 identical; pre-commit (fprettify-stable), Sphinx-W.