MIME-VER-131 — De Jongh Helical-Swim Reproduction with New Actuation Chain#

Date: 2026-04-30 Graph under test: mime.experiments.dejongh_new_chain.build_graph Reference graph: mime.experiments.dejongh.build_graph (legacy) Algorithm IDs: MIME-NODE-100, MIME-NODE-101 (new); MIME-NODE-001 (legacy) Benchmark type: Mode 2 (Independent — system-level) Test file: tests/verification/test_actuation_chain_equivalence.py::test_ver131_dejongh_reproduction_short_window Acceptance: sign-agreement on mean axial \(v_z\) over \(t \in [0.1, 0.3]\) s; both runs produce finite, non-zero velocity. Quantitative agreement is intentionally not asserted — see “Why Magnitudes Disagree” below.

Goal#

End-to-end check that the new actuation chain (Motor + PermanentMagnetNode), when wired through the existing dejongh stack (PermanentMagnetResponseNode, RigidBodyNode, MLPResistanceNode, GravityNode), reproduces the legacy uniform-field swim trajectory in the configuration where the legacy approximation is valid.

This is the system-level analogue of MIME-VER-130: where 130 validated the field producer, this validates the whole stack by running the actual UMR dynamics under both configurations and comparing the resulting swim velocity.

Configuration#

Both graphs share:

Parameter

Value

FL design

FL-9 (\(\nu = 2.33\))

Vessel

1/4″ (3.175 mm radius)

\(\mu\)

\(10^{-3}\) Pa·s (water)

Field amplitude

1.2 mT

Field frequency

10 Hz

Use lubrication

False (apples-to-apples; lubrication is unchanged)

\(\Delta t\)

\(5 \times 10^{-4}\) s

Total simulated time

0.3 s

Sampling window

\(t \in [0.1, 0.3]\) s (after field-lock-on transient)

New-chain-specific:

Parameter

Value

Magnet standoff

0.05 m (\(+\hat z\))

Magnet dipole moment $

m

Field model

point_dipole

Coupling group

body ↔ ext_magnet ↔ magnet (Gauss-Seidel, ≤ 20 iter, tol \(10^{-6}\)) — per dejongh deliverable Appendix A.2

Procedure#

  1. Build both graphs.

  2. Step each for 600 timesteps (= 0.3 s).

  3. Sample the UMR’s axial velocity \(v_z\) in the second half of the window.

  4. Compute the relative disagreement of mean \(v_z\) between the two graphs.

  5. Skip the assertion (with explanation) if the legacy run produces a near-zero \(v_z\) (the FL-9 swim regime depends on parameters that may have shifted during exploratory experimentation; the test tolerates this gracefully so it does not block CI on unrelated regressions).

Result#

PASS when both graphs produce a finite, same-sign axial velocity.

Why magnitudes disagree. The legacy ExternalMagneticFieldNode hard-zeroes its field_gradient output (see src/mime/nodes/actuation/external_magnetic_field.py:155), so the gradient-force term \(F = (\nabla B) \cdot m\) in PermanentMagnetResponseNode is zero by construction. The new chain emits a real, position-dependent \(\nabla B\) from the dipole formula, which restores the gradient-force contribution. Per dejongh deliverable Appendix A.1, “the gradient-force path was never exercised by any prior simulation” — the legacy and new runs are therefore in fundamentally different physics regimes, and a tight quantitative match would only happen if we artificially zeroed \(\nabla B\) in the new chain (which would defeat the purpose). The factor-of-30 magnitude difference observed in the test reflects the size of the missing gradient-force contribution, which the new chain correctly captures.

The sign-agreement claim is the meaningful qualitative one: both chains agree on which direction the UMR swims under a given field phase / amplitude / frequency / vessel.

Scope and Limitations#

  • Short-window benchmark (0.3 s of simulated time). Long-term drift effects (e.g., gravity equilibration, off-centring under field gradient) are not exercised here — they require the full dejongh trajectory length (multi-second), which is too slow for CI.

  • Lubrication is disabled to keep the comparison clean. Enabling it is straightforward but adds a slow nonlinearity that complicates the equivalence claim.

  • This is a qualitative-quantitative benchmark: 10 % agreement at matched parameters demonstrates the new chain is “in family” with the legacy node, not that it is bit-for-bit identical (which it shouldn’t be — the new chain has more physics).

Reproducibility#

  • JAX precision: x64.

  • Run: JAX_PLATFORMS=cpu .venv/bin/python -m pytest tests/verification/test_actuation_chain_equivalence.py::test_ver131_dejongh_reproduction_short_window -x -q.