Skip to content

P92 life-data screen: fixed calibration and bounded holdout

Sources and scope

Zhang et al., Metals 10 (2020) 307, Table 2, reports eight life observations at two temperatures, two total axial strain ranges and three tensile dwell durations. Section 2 defines failure by the authors' 10% peak-stress-drop criterion. These are not silently relabeled rupture times, crack-size lives, or allowable component lives. Source values, heat treatment, loading rate and split are retained in evidence/creep/p92-life-table.json.

The separate BAM/Zenodo data contain service-like cyclic deformation histories at 620 C. They are a different experiment and are not pooled with this table as though heat/batch, loading and failure definitions were identical.

ASTM E2714's public scope distinguishes specimen creep-fatigue deformation/crack formation from full-component testing, and recommends complementary continuous-cycling and creep data for the same material and temperatures. This screen does not implement the paid standard or claim code compliance.

One deliberately simple baseline

Before numerical evaluation, use IDs 1,3,4,5,7,8 for calibration and 2,6 for within-source validation. Fit only the training rows in log-life coordinates:

ln(N) = a + b*ln(total_strain_range/.008)
          + c*ln(1+tensile_dwell_seconds/60) + d*(T_C-600)/50
b, c, d <= 0

This is an explicitly chosen empirical screening ansatz, not a published constitutive law or a calibrated damage mechanism. The four parameters have six fitting observations. Held-out cases share the low strain range and an intermediate dwell: they test interpolation within this source, not a new material batch or an unseen strain regime. No alternative model shopping, resplitting or refitting to the holdout is used.

Reproduce with python tools/p92_life_screen.py. The script prints complete JSON with source, split, coefficients, every prediction, ratios and limitations. 14 tests verify source transcription, fixed baseline, no holdout-label leakage, monotonicity, rank admission and rejection outside the declared input domain. The regression's numerical expectations do not constitute scientific acceptance thresholds selected before the experiment.

Result

The two held-out observed lives are 1190 and 853 cycles; predicted lives are 961.892 and 788.484 cycles, respectively. Signed relative errors are -19.17% and -7.56%; multiplicative error factors are 1.237 and 1.082. Calibration predictions and their residuals remain in the report rather than being combined with validation accuracy. This establishes a reproducible, modest within-source prediction baseline, not an uncertainty bound or engineering qualification.

This is already a combined creep-fatigue life relation. Feeding it as a pure-fatigue curve into a time-fraction-plus-fatigue damage sum would count the dwell effect twice. It is intentionally not wired into the production assessment API as a P92 material card.

Stop conditions and next admission gates

  • Keep the baseline fixed; further fit iterations on these eight rows cannot create independent validation or repeat-specimen scatter.
  • Admit only 600 or 650 C, total strain range .008-.012, tensile dwell 60-600 s, and the recorded loading rate/environment. No temperature interpolation, zero-dwell prediction, compressive dwell, weld or multiaxial claim.
  • Obtain compatible pure-fatigue and creep-rupture/ductility data if separating damage mechanisms. Record heat treatment, specimen, failure criterion and units before combining sources; do not borrow missing parameters by name.
  • Test additional independently selected lives, then state prediction scatter and domain. Engineering allowable life needs a justified conservative margin, not merely a factor-of-two scatter plot.
  • Co-located structural histories are a separate data prerequisite. Total strain cannot automatically replace the mechanical strain required by a strain-life assessment. A material response model alone has no fracture clock.

Status: research baseline calibrated and held-out evaluated; engineering life qualification remains open with explicit data requirements.

Cross-publication audit

Wang et al., Metals 9 (2019) 183, Table 1 (page 4, visually checked), lists CF600S04T01/T03/T10 with lives 1303/1190/675. At 600 C, amplitude .004 and tensile dwells 1/3/10 minutes, these exactly match rows 1/2/3 after amplitude-to-range and minute-to-second conversion. Publication independence is not specimen independence: exclude these observations from external validation pending specimen provenance. This includes the current held-out row 2; it remains a within-source holdout, not external evidence. The JSON report exposes this overlap warning.

The same table also contains zero-dwell repetitions, compressive holds and a 60-minute tensile hold. Those are useful future mechanism/domain tests, but are outside this baseline's admitted loading domain. Do not extend the domain just to obtain additional validation rows. No coefficients were refitted.

Published pure-fatigue interface check

tools/p92_pure_fatigue_screen.py reuses the public CoffinMansonBasquinCurve with the 2019 paper's Table 2 coefficients at 600/650 C and Table 1 no-dwell observations. All replicate lives remain separate. These observations contributed to the authors' fit: this is a reproduction check, not a new holdout. The table's failure criterion has not been independently resolved and is not borrowed from the 2020 paper.

For this screen only, admissible life is bounded by observed life extrema at each temperature. These are our conservative numerical admission limits, not published design allowables. The .002/.008 amplitude endpoint queries fall outside the corresponding fitted amplitude bounds and are explicitly reported as rejected; they are not silently extrapolated or discarded. This does not mean those experimental points or the published formula are invalid. Interior .004-amplitude predictions are 1800.22 and 1090.80 cycles. At 600 C the four observed repetitions span 1426-2186 cycles; a single curve does not eliminate this scatter. Three tests cover the published coefficients, replicates, domain decisions and direct-equation roundtrip.

Together these screens supply a traceable starting dataset and demonstrate existing software interfaces. They do not yet supply a jointly validated creep-rupture curve, interaction envelope or component-life material card.