P92 full-path diagnostic and next material gate¶
Decision¶
The first-cycle loading/holding/reversal path now runs through the ordinary
creep_solid -> model.step -> SimulationResult lifecycle without resetting
stress at the start of a hold. The frozen high-stress Norton baseline does
not reproduce the low-stress SLR hold. Do not promote it as a validated
P92 material or expand the core material catalogue to conceal that limitation.
This is a scientific discriminator, not a new FEM solver or a release gate
that indefinitely blocks already verified engineering improvements.
Sources and protocol¶
- BAM/Zenodo 7198219 v1, Sonntag et al. (2022), CC BY 4.0: checksum-pinned digital RF and SLR records.
- Data in Brief article: Tables 1 and 2 define the nominal strain paths and physical-time segments.
- Related experimental paper, open manuscript: pages 8–9, Figure 10 and Table 2 were visually inspected. Its mechanism-map normalization uses E=151 GPa; the paper discusses competing creep/recovery and possible changes of dominant mechanism during long low-stress relaxation. This is a mechanism hypothesis, not a calibrated constitutive equation or proof of a unique cause.
tools/validate_p92_loading_path.py preserves the previous RF first-hold fit
(r200=1.590673 MPa/s, n=9.687639). It estimates an effective initial slope from
the predeclared RF 0.15–0.60 s window: E=146259.633 MPa. Two neighboring
windows and the paper's 151 GPa value are sensitivity checks, never selected
by their SLR prediction errors. This slope is not an independent modulus test.
For each family, use all first-cycle measured strain samples, subtract only the first measured strain offset, start with zero stress/internal strain, and advance continuously. No measured stress initializes or resets the full-path prediction. Separately run the nominal published strain path, without fitted time shifts or peak alignment. Comparison ends at available path coverage; nominal endpoints and actual recorded endpoints differ slightly.
The experiment is isothermal at 893.15 K; measured temperature variation is reported but not modeled. Nominal experimental stress is approximated by infinitesimal axial Cauchy stress. It is a homogeneous free-lateral unit cube, not a reconstructed specimen, spatial convergence study or lifetime model. The assumed initial stress is reported alongside the near-zero measurement.
Executed evidence¶
An independent scalar Radau integration of
dσ/dt = E*dε/dt - r200*sign(σ)*(abs(σ)/200)^n resolves every input knot.
The FEM uses the same prescribed piecewise-linear strain, but its existing
backward-Euler material integration, global equilibrium and state owner.
Every original input interval is subdivided by 1, 2 and 4; no input corner is
skipped. The separate scalar implementation is an oracle, not a Step backend.
| Path | FEM increments | Maximum FEM–oracle error, MPa |
|---|---|---|
| RF | 474 / 948 / 1896 | 2.20336 / 1.12342 / 0.56750 |
| SLR | 925 / 1850 / 3700 | 2.19469 / 1.12071 / 0.56664 |
Both numerical gates pass: decreasing errors, finest error below 1 MPa, lateral stress below 1e-6 MPa, prescribed-strain error below 1e-10. Ratios approximately 1.96–1.98 agree with first-order temporal behavior. These are predeclared numerical criteria, not experimental material tolerances.
| Continuous baseline vs experiment | RF | SLR |
|---|---|---|
| Measured-input whole-cycle time-weighted RMSE, MPa | 3.73952 | 67.06971 |
| Measured-input maximum error, MPa | 40.00434 | 82.56211 |
| Nominal-input whole-cycle time-weighted RMSE, MPa | 4.64323 | 67.22317 |
SLR's 184–1984 s nominal zero-strain interval dominates the discrepancy: RMSE 73.55028 MPa. Varying E across the inspected range leaves this error at 73.39–75.74 MPa; using measured rather than nominal input does not remove it. For a post-hoc discriminator only, initialize the separate analytical hold solution from the measured 84.41 MPa at 194.008 s. It predicts 83.76794 MPa at 1978.331 s, versus 45.3 MPa observed. That conditional test is not a state reset inside the full-path simulation or a new independent validation.
The evidence rules out time refinement alone, the inspected E range, and nominal-versus-measured strain input as sufficient fixes for this baseline. It does not uniquely separate transient creep, static recovery, backstress, temperature sensitivity or specimen/measurement uncertainty. A better law must explain both stress regimes and reversal, not merely lower total RMSE.
Evidence: evidence/creep/2026-11-01-p92-full-loading-path.json, with source
hashes, script hashes, versions, fit separation, phase errors and sensitivities.
The combined protocol/history/analytical regression passed 39 tests in
34.13 s. Seven new offline tests cover input rejection, exact linear ramp/
hold/reversal, RF-only slope selection, nominal knots, FEM refinement, reaction
agreement and checkpoint continuation at reversal. No network data is needed
by these tests. Experimental acceptance remains unassigned/false.
The documentation/architecture suite passed 44 tests in 3.94 s; Ruff,
generated-document freshness, strict MkDocs build and frozen-tool hash checks
also passed. The initial architecture-check command named a nonexistent test
file and ran no tests; the corrected command produced the 44-test result.
Reproduce¶
In the scientific environment, using the checkout/package under test:
python tools/validate_p92_loading_path.py --rf /path/to/Data_RF.lis \
--slr /path/to/Data_SLR.lis --fem --output /path/to/new-report.json
No automatic download, environment modification or report overwrite occurs. Exit status checks numerical verification; experimental status remains an explicit separate field, never inferred from successful computation.
Resume plan¶
- Integration update (2026-11-05): #138 is merged, as are #139–#141. Continue scientific gates separately from the completed engineering slices.
- Keep this Norton baseline frozen as a negative transfer benchmark. Screen the existing Sinh and creep/plastic assets for representational sufficiency before adding a new core law. Any recovery/softening proposal needs state, tangent, dissipation conventions and independent tests first.
- SLR is now used for diagnosis/model selection. Do not call later retuned predictions on it blind validation. Freeze a new compatible condition or explicitly distinguish same-specimen later-cycle tests from independent specimen validation. SLC has stress control and cannot reuse a strain-held boundary script unchanged.
- Preserve same-point stress/strain/temperature and restart identity for life assessment, but do not attach a P92 rupture/life curve without source and calibration-domain evidence. Temperature-dependent UserMaterial, regional joint plasticity, contact and MPI remain separately bounded work items.
No new default solver settings, installed environment, private material, customer files, package build or composite long run changed in this sprint.