P92: response, calibration and life evidence¶
AgentFEM now has a reproducible P92 research baseline and reusable history/life assessment infrastructure. The remaining engineering task is to establish prediction on compatible independent material data and define the intended component assessment. More solver iterations cannot substitute for that data.
A coherent-cohort source audit now includes visually checked creep, fatigue and dwell observations. Fixed printed creep coefficients have unresolved strain-scale discrepancies; the data are not an installed material card or a qualified coupled-life model.
What can be reproduced now¶
The specimen life workflow admits the 18 DB source rows as 18 exact-condition groups. It retains failures and provenance without inventing run-outs, replicate scatter or P92 confidence-qualified design lives.
These are repository research tools, not installed material cards. Run them from a source checkout with AgentFEM and its scientific dependencies installed. Each command emits JSON, keeps source identity and states its validation role.
python tools/p92_life_screen.py
python tools/p92_pure_fatigue_screen.py
python tools/p92_rupture_screen.py
python tools/p92_prior_damage_audit.py
python tools/p92_rate_fatigue_screen.py
python tools/p92_db_cohort_screen.py
The cyclic figure screen additionally needs an existing continuous-reference
campaign: python tools/p92_cyclic_figure_screen.py campaign.json. Its
reproduction record explains generation
and reuse; do not start a long campaign just to inspect the archived conclusion.
| Question | Evidence and result | Meaning |
|---|---|---|
| Can a bounded combined-life model be calibrated without fitting its holdout? | Six training observations, two fixed held-out observations; predictions 961.892 and 788.484 cycles versus 1190 and 853 | Within-source interpolation baseline, not an independent batch or allowable |
| Do the published response equations reproduce cyclic observations? | Fixed-parameter first- and long-cycle experimental figure screens retain reading sensitivity and residual discrepancies | Response evidence; an observed failure count used to place a plotted marker is not predicted failure |
| Can supplied rupture observations enter a reusable assessment? | Four completed rupture tests, split into unaged/aged conditions; exact node recovery and half-life time fraction 0.5 | Numerical interface verification, not validation of interpolation or extrapolation |
| Can a published pure-fatigue relation use the fatigue API? | Separate source-defined strain-life parameters and tests | Reproduction, not a new independent fit and not permission to combine incompatible sources |
The source audits and full numerical records are in the
life screen,
cyclic screen, and
external-source admission audit.
The generic assessments.TabulatedRuptureCurve retains stress/time units,
stress measure, material condition, temperature and failure definition. It
interpolates only within its declared stress interval at its supplied
temperature; it does not select a P92 curve for the user.
The prior-damage audit adds a traceable remaining-creep-life dataset and audits published predictions, including their large high-damage errors. It is not a newly executed material model. This offers a concrete next reproduction route without pooling unrelated P92 conditions.
The equation reference now executes the bounded as-received creep law and agrees with an independent time-domain integration. The published backstress equations are inconsistent, and the dimensionally consistent reading still differs from the tabulated prediction by approximately 3.67%. This is numerical research evidence, not a promoted P92 material card or engineering lifetime validation.
Two life routes, not one mixed formula¶
The service-aged rate screen adds a separate 600 C cohort and reproduces published fatigue coefficients without refitting. Cross-rate errors remain visible (up to +32.56% among the cited admitted high-to-low-rate comparisons); no blind validation is claimed.
Combined test-life route: a fitted creep-fatigue life relation predicts the source-defined combined test outcome for an admitted waveform. Do not feed it into a pure-fatigue damage term and add creep damage again.
Separated assessment route: compatible pure-fatigue and creep-rupture or ductility information supports separately defined damage measures. The chosen interaction envelope, stress/strain measures and design margins must come from the assessment being followed. A linear sum is a modeling choice, not a universal validated P92 rule.
The public scope of ASTM E2714-26 recommends complementary continuous-cycling and creep data for the same material and temperatures. Its crack-formation outcome is not automatically the published P92 paper's peak-stress-drop criterion. This reference guides data selection; it is not a claim of compliance with the full standard.
Smallest next engineering validation package¶
Use the engineering-life uncertainty guide to distinguish population scatter, parameter/model uncertainty, predictive intervals and design margins. The target is a decision with explicit uncertainty, not exact agreement with every specimen or a probability band around an unqualified model. The guide records reusable methods and proposed next steps; it does not add a statistical-life API.
- Choose the application and failure outcome: parent metal or weld, temperature range, loading rate, tensile/compressive holds, environment and whether the target is specimen failure, crack initiation, or a code-based component margin.
- Supply specimen-level observations with heat/product/heat-treatment identity, units and stress/strain conventions. Retain interrupted/runout tests as such; do not replace them with rupture events. Repeated tests quantify scatter.
- Freeze a calibration subset and model choices before evaluating independent observations. Check for reused specimens across publications. The current eight-row table cannot create an independent validation set by being split differently or cited from a second paper.
- Pass numerical checks with those frozen parameters: material point, homogeneous structure, time refinement, failed-trial rollback and restart. Use co-located accepted stress/strain/temperature histories, not extrema from different locations. Then assess a nonuniform component with spatial/time refinement and the actual boundary conditions.
- Report prediction error and scatter separately from design margin. Select application-specific acceptance criteria before inspecting validation errors; do not promote an arbitrary factor-of-two plot to an engineering guarantee.
NIMS sheet 48B is a valuable next source: it includes several T92/P92 heats and long-duration rupture observations. Access and use are through the NIMS data service; the announcement alone does not supply the numerical curves. Different heats are not silently pooled.
Compatible does not mean necessarily the same heat¶
Same-heat tests are useful, not an absolute software admission rule. A cross-heat research assessment may proceed with explicit transfer assumptions, stratified validation and reported scatter. It must not be labelled same-heat validation or a component allowable. Preserve individual source identities instead of rewriting their material-condition labels to make a combination fit.
The public EPRI report 1019778 (2010), Table 5-1 and sections 5.2–5.3, calls for pedigree/distribution/scatter review and distinguishes deformation interaction from damage interaction. Our cross-heat research policy is an implementation decision informed by that guidance, not a claim that the report supplies P92 design curves.
Integrated software foundation¶
The nonuniform bending screen adds four mesh levels and two time grids using the ordinary structural workflow. Total-reaction screening and a small-model restart pass; relaxation drop and clamp hotspot spatial convergence remain open. This separates a useful numerical composition test from experimental multiaxial validation.
Small-strain selected histories and prescribed temperature, structural cyclic checks, and thermal clock/restart safeguards are merged through PRs #152, #155, #158 and #160–#163. The combined installed-use acceptance includes restart comparison and real VTK reading. The structural cycle record includes 435 imposed cycles at 650 °C with time refinement, not a prediction of failure. No release or engineering qualification follows automatically from merging these interfaces.
Stop retuning this baseline when the missing information is experimental. Continue useful software work on reusable material/state/result contracts; resume engineering calibration when the selected application and compatible validation observations are available.