The regression suite covers pinned IMRv2 trajectories across radial equations, forcing, vapor, heat transfer, mass transfer, and the specialized constitutive models. It also checks:
- composable neo-Hookean/Newtonian dynamics against the closed-form Kelvin-Voigt path;
- elastic and viscous reduction limits;
- analytic stress-rate and acceleration tangents against centered finite differences;
- Giesekus and linear PTT convergence to Oldroyd-B;
- sparse coupled thermal-memory integration;
- mechanical, thermal, distributed, and collapse-shooting sensitivities against independent centered differences;
- likelihood Jacobians and retained multistart endpoints.
Two statistics are reported per pinned trajectory, because they measure different things (issue #23).
The pointwise maximum sits at a collapse in every pinned case. There
|dR/dt| ~ 3.3e5 /s, so it is dominated by a sub-nanosecond timing difference
rather than by radius accuracy: shifting our own solution by 25 ps removes
about 77% of it, taking Keller-Miksis neo-Hookean from 8.6e-06 to 2.0e-06.
The residual matches the deviation away from collapses, 2-10e-06, so the true
pointwise agreement is several times better than the maxima below suggest. The
maximum cannot be tightened without measuring integrator phase.
The median carries no such sensitivity, and is what the suite bounds
tightly. It is far more responsive to real error: a 3e-5 relative
perturbation of the polytropic pressure moves Keller-Miksis NHKV's median by a
factor of 38 and its maximum by only 1.35x. Bounds are per case rather than
uniform, because baseline medians span 3e-08 to 1.6e-06 and one threshold
would be set by the worst case.
Representative maximum absolute radius-ratio deviations from pinned IMRv2 trajectories are:
| case | maximum deviation |
|---|---|
| Zener, Deborah number 2, stretch 6 | 2.6e-05 |
| neo-Hookean Kelvin-Voigt parameter grid | 6.3e-05 |
| quadratic Kelvin-Voigt | 1.7e-05 |
| Oldroyd-B | 6.2e-05 |
| thermal and mass-transfer branches | 1.6e-05 |
| compressible radial-equation families | 1.6e-05 |
These are numerical comparisons with the reference implementation, not estimates of physical-model error.