Skip to content

fix: a finite-EI dynamic step on the inertia's round-off floor converges - #168

Merged
SMI-Lab-Inha merged 1 commit into
devfrom
fix/viv-tolerance-floor
Nov 15, 2026
Merged

SMI-Lab-Inha merged 1 commit into
devfrom
fix/viv-tolerance-floor

Conversation

@SMI-Lab-Inha

Copy link
Copy Markdown
Owner

Summary

A finite-EI cable could fail its first dynamic step at a small time step: Newton stagnated just above the tolerance, and the step recovery's smaller substeps only made it worse. The fix is a stall test on the finite-EI route, like the one the EI = 0 route already has. A step that meets the tolerance is unchanged; every example and test deck gives bit-identical output.

How it was found

The Delta Flume riser (64 elements, 13 m, at rest in a 0.95 m/s current, VIV lift on) at $\Delta t_M = 1.25\times10^{-4}$ s. The residual stopped at 1.06 times the tolerance, and every 4-fold smaller substep raised the stagnant residual about 16-fold:

Substeps 1 4 16 64 256 1024
Stagnant residual (N) 3.8e-8 6.2e-7 8.6e-6 1.3e-4 2.6e-3 4.7e-2

Cause

The 16-fold growth per 4-fold step is the $1/\Delta t^2$ of the inertia, not the $1/\Delta t$ of a velocity-coupled force such as drag or the VIV lift. The inertia takes the acceleration $(q - q_n)/(\beta\Delta t^2) - \dots$ from coordinates stored to their last bits, so its residual cannot be settled below

$$ r_\text{floor} \approx \varepsilon \sum_j |M_{ij}|,|q_j| ,/, (\beta \Delta t^2). $$

  • The rows that stalled were the axial rows of the nodes 8.5 to 12 m deep, not the cross-flow rows the lift loads: the Newton correction they needed (a few 1e-16 m) is below the 1.8e-15 m spacing of those coordinates.
  • Without the lift the step starts at the static equilibrium and converges at its predictor. With it, the line near rest sets no inertial force scale, so tol × force scale (here the axial round-off floor) is below the inertia's floor.
  • A floor on the force scale built from that bound was rejected. Where it binds it would accept residuals up to 16 times the bound, some 60 times what Newton reaches (1.7e-7 against 3.8e-8 N on the riser).

Fix

The iterate is accepted as converged when all of these hold:

  1. no step length of a fresh Newton correction reduces the residual;
  2. the residual is below $10^{-4}$ of the force scale;
  3. the correction moves no free coordinate by more than $16\varepsilon$ times its scale. The scale is the largest coordinate magnitude or the element length for positions, and 1 or the largest tangent or twist magnitude for those DOFs.

Results

  • Riser: now runs at $\Delta t_M = 1.25\times10^{-4}$ s. Over 2 s its node positions and fairlead tension continue the sequence of $5\times10^{-4}$ and $2.5\times10^{-4}$ s:

    Step pair Node position difference Fairlead tension difference
    $5\times10^{-4}$ s vs $2.5\times10^{-4}$ s 1.3 mm 5.5 N
    $2.5\times10^{-4}$ s vs $1.25\times10^{-4}$ s 0.55 mm 2.9 N
  • New test test_hermite_viv (round-off floor): the test's 100 m line hung vertically from 50 to 150 m depth, from rest with the lift at $\Delta t = 2.5\times10^{-5}$ s.

    • Before: the first step stagnated at 2.4 times the tolerance through its 50-iteration budget and failed.
    • After: each step stops on the floor after two iterations, and the cross-flow displacement at 0.02 s matches the same line at $\Delta t = 10^{-3}$ s within 1 %.

Checks

  • Bit-identical outputs on the old and new driver: 229 files over 118 example and test-data decks.
  • C API: five coupled steps of every deck give bit-identical line arrays (1685 arrays over 70 decks).
  • Full test suite on the default and the Prescott OpenBLAS kernel (424 tests); the 66 example decks run with no fallback note. After the rebase onto dev 6c469fe: 426/426 on the default kernel.

A finite-EI cable could fail its first dynamic step at a small time
step: Newton stagnated just above the tolerance, and the step
recovery's smaller steps only raised it. Found on a 13 m riser (64
elements) at rest in a 0.95 m/s current taking the VIV lift at dtM
1.25e-4 s: the residual stopped at 1.06 times the tolerance, and each
4-fold smaller substep raised the stagnant residual about 16-fold
(3.8e-8, 6.2e-7, 8.6e-6, 1.3e-4, 2.6e-3, 4.7e-2 N), to 3.4e4 times the
tolerance at 1024 substeps.

The 16-fold growth per 4-fold step is the inertia's 1/dt^2, not the
1/dt of a velocity-coupled force (drag, VIV lift): the inertia takes
the acceleration (q - q_n)/(beta dt^2) - ... from coordinates stored
to their last bits, so its residual cannot be settled below about eps
sum_j |M_ij| |q_j| / (beta dt^2). The rows that stalled were the axial
rows of nodes 8.5 to 12 m deep, not the cross-flow rows the lift
loads: the Newton correction they needed, a few 1e-16 m, is below the
1.8e-15 m spacing of those coordinates. Without the lift the step
starts at the static equilibrium and converges at its predictor; with
it the line near rest sets no inertial force scale, and tol times the
force scale (here its axial round-off floor) lies below the inertia's
floor. A floor of the force scale built from that bound, where it
binds, would accept residuals up to 16 times the bound, some 60 times
the floor Newton reaches (the bound sits about 4 times above it:
1.7e-7 against 3.8e-8 N on the riser, 4.7e-3 against 1.2e-3 N in the
test below). The fix is therefore a stall test, as on the EI = 0
route: when no step length of a fresh Newton correction reduces the
residual, the residual is below 1e-4 of the force scale, and the
correction moves no free coordinate by more than 16 eps times its
scale (the largest coordinate magnitude or element length; 1 or the
largest tangent or twist magnitude for those DOFs), the iterate is
converged.

The riser now runs at dtM 1.25e-4 s, and over 2 s its node positions
and fairlead tension continue the time-step sequence of dtM 5e-4 and
2.5e-4 s (differences 1.3 mm and 0.55 mm, 5.5 N and 2.9 N). A step
that meets the tolerance is unchanged: every example and test-data
deck gives bit-identical outputs on the old and new driver (229 files
over 118 decks), and through the C API five coupled steps of every
deck give bit-identical line arrays (1685 arrays over 70 decks), and
the full test suite (424 tests) passes on the default and the Prescott
OpenBLAS kernel, where the 66 example decks also run with no fallback
note.

Test: test_hermite_viv round-off floor, the test's 100 m line hung
vertically from 50 to 150 m depth, from rest with the lift at dt
2.5e-5 s. Before, the first step stagnated at 2.4 times the tolerance
through its 50-iteration budget and failed; now each step stops on the
floor after two iterations, and the cross-flow displacement at 0.02 s
matches the same line's at dt 1e-3 s within 1 % (to 5 digits).

Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
@SMI-Lab-Inha

Copy link
Copy Markdown
Owner Author

@codex review

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 15, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-15T20:07:44.002962Z 48f5b4d Manual request
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@read-the-docs-community

Copy link
Copy Markdown

Documentation build overview

📚 CableDyn | 🛠️ Build #35911271 | 📁 Comparing 48f5b4d against latest (789cb6d)

  🔍 Preview build  

14 files changed · ± 14 modified

± Modified

@chatgpt-codex-connector

Copy link
Copy Markdown

Codex Review: Didn't find any major issues. 🚀

Reviewed commit: 48f5b4d392

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

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".

@SMI-Lab-Inha
SMI-Lab-Inha merged commit efbfbba into dev Nov 15, 2026
16 checks passed
@SMI-Lab-Inha
SMI-Lab-Inha deleted the fix/viv-tolerance-floor branch November 15, 2026 22:10
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant