EUVIMEDEUVIMED entdecken
Crossref

Confinement Transfer, Training-Only Model Selection and Bounded Sensitivity of Rock-Strength Surrogates

Emir Hadzic

Geotechnical Science and Engineering Journal · 05.08.2023 · 2023 (1) · 164-194

DOI: 10.68304/gsej2023.a108 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Geotechnical Science and Engineering Journal, online erschienen am 05.08.2023. Verfasst von Emir Hadzic.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The numerical reliability of a rock-strength surrogate depends on the omitted stress conditions and the scale used to aggregate prediction errors. The analysis examines sixty-seven true-triaxial failure states for sandstone, limestone and trachyte. An affine surface and a restricted quadratic surface are compared through record deletion and minimum-confinement-level deletion. A full second-order polynomial and a constrained Mohr--Coulomb compression relation provide sensitivity comparisons. Direct refitting is checked against exact block-deletion identities, while retained stress hulls distinguish geometrically supported predictions from extrapolation. For the restricted quadratic, confinement-deletion root-mean-square errors are 7.12, 29.53 and 11.31 MPa for sandstone, limestone and trachyte, compared with 6.96, 21.12 and 22.94 MPa for the affine surface. Their pooled dimensional errors are nearly identical, 19.70 and 19.73 MPa, because limestone deterioration offsets trachyte improvement. However, the quadratic reduces pooled error normalized by material UCS by 30.30%, and the mean material RMSE by 6.00\%. The full polynomial improves sandstone transfer but increases pooled dimensional confinement-deletion error to 39.09 MPa. A nested confinement-selection procedure gives a pooled held-out RMSE of 16.71 MPa. Exact componentwise perturbation propagation identifies a maximum limestone gain of 14.92 for the full polynomial. These results distinguish material-dependent model comparison, training-only selection and sensitivity to bounded strength errors. Stress-domain checks and explicit coefficient calculations show how sparse branch support can destabilize flexible surrogates.

Zitieren

APA

Hadzic, E. (2023). Confinement Transfer, Training-Only Model Selection and Bounded Sensitivity of Rock-Strength Surrogates. Geotechnical Science and Engineering Journal, 2023(1), 164-194. https://doi.org/10.68304/gsej2023.a108

Vancouver

Hadzic E. Confinement Transfer, Training-Only Model Selection and Bounded Sensitivity of Rock-Strength Surrogates. Geotechnical Science and Engineering Journal. 2023;2023(1):164-194. doi:10.68304/gsej2023.a108

BibTeX
@article{Hadzic2023,
  author = {Hadzic, Emir},
  title = {Confinement Transfer, Training-Only Model Selection and Bounded Sensitivity of Rock-Strength Surrogates},
  journal = {Geotechnical Science and Engineering Journal},
  year = {2023},
  volume = {2023},
  number = {1},
  pages = {164--194},
  doi = {10.68304/gsej2023.a108}
}
Kostenlos starten →