P92: first executed public-data relaxation comparison¶
Why this source¶
The BAM dataset, Zenodo 7198219 v1 by Sonntag, Jürgens, Uhlemann, Skrotzki and Olbricht (2022), CC BY 4.0, supplies digital time/stress/strain/temperature records, not just life tables. The associated Data in Brief article explains nominal stress, percent strain, Celsius and sparse cycle recording. This is P92 at nominal 620 °C; it is not Alloy 617 and does not replace the separate Alloy 617 source audit. No private applicant or customer data is used.
Fixed, deliberately small experiment¶
tools/validate_p92_relaxation.py reads checksum-pinned original files. It
converts percent to strain and Celsius to kelvin without resampling or filling
missing cycles. Select cycle 1, raw time 8–180 s, avoiding ramp/servo settling
and reversal; retain the actual first/last samples. RF tensile holding fits
the effective law
Trapezoidal time weights prevent dense early sampling from dominating the fit. Three starting points converge to the same least-squares cost. Freeze the parameters before predicting the separate SLR compressive hold. The measured initial stress is an explicit initial condition in each hold; it is not a predicted peak. No held-out curve values enter optimization.
This identifies E*A and n, not E and A separately. Full-cycle plasticity,
cyclic softening, damage, thermal sensitivity, initial loading and lifetime
remain outside the fit. Measured strain/temperature drift is reported but
neglected by this ideal constant-strain/isothermal baseline. Experimental
nominal stress is approximated by small-strain axial stress, explicitly.
Executed evidence¶
The frozen fit gives r200 = 1.590673 MPa/s, n = 9.687639, with no active
optimization bound. Time-weighted stress RMSE is 0.502629 MPa for RF and
6.496178 MPa for the held-out SLR hold. Maximum SLR error is 8.328164 MPa;
its endpoint prediction is −149.954699 MPa, against −142.85 MPa measured.
These are descriptive errors, not a passed experimental acceptance criterion.
No uncertainty model or independently justified material tolerance has yet
been established. Do not retune against this holdout and still call it held out.
The ordinary creep_solid → model.step workflow realizes the same conditional
law on a free-lateral unit cube. E=100000 MPa is a numerical realization,
not a measured P92 modulus; set A=r200/E and initial strain=σ0/E. An
offline test doubles E while preserving E*A and verifies the same stress.
No new solver, P92-specific core material, or project Newton loop is added.
| Increments | Max FEM–continuous time error (MPa) | Endpoint error (MPa) |
|---|---|---|
| 32 | 2.160319 | −0.805955 |
| 64 | 1.162444 | −0.411316 |
| 128 | 0.606418 | −0.208049 |
This is temporal verification of homogeneous equilibrium, not spatial
convergence or reproduction of a physical specimen. Raw point stress agrees
with the strong-boundary reaction; restart retains the accepted point history.
Eight offline tests cover parsing, units, sparse cycles, source identity,
time weighting, signed analytical response, synthetic parameter recovery and
actual FEM pause/restart. They do not require network or the experimental files.
The combined point-history, experiment-protocol and CI/package checks passed
110 tests in 27.48 s. The executable numerical gate passed with error
ratios 1.8584 and 1.9169. Full machine-readable evidence is retained at
evidence/creep/2026-11-01-p92-conditional-relaxation.json in the repository.
Reproduce¶
Obtain Data_RF.lis and Data_SLR.lis from the versioned Zenodo record. Keep
them outside the repository; no download or package install occurs in the tool.
In the normal scientific environment, from a source checkout using the package
under test:
python tools/validate_p92_relaxation.py --rf /path/to/Data_RF.lis \
--slr /path/to/Data_SLR.lis --fem --output /path/to/new-report.json
Existing reports are never overwritten. The report records source SHA-256, tool SHA-256, versions, actual sample endpoints, fit/holdout errors and separate numerical evidence. A failed numerical gate retains its report and exits nonzero; a descriptive experimental comparison never promotes material trust.
Next scientific decision¶
Use this baseline to locate what is missing, not to set a product promise. Before full-cycle validation, acquire/estimate E from an independently reviewed elastic segment, retain the full controlled loading history, choose and verify a law that admits cyclic softening, and freeze a new condition/specimen holdout. Then integrate through the existing material/Step/state contract. The longer zero-strain SLR hold and stress-controlled SLC route are not silently treated as the same strain-held experiment. A generic user-material temperature contract remains a separate implementation task.