Scholarlink

<p>This study develops a reliability-aware machine-learning framework for compressive-strength prediction and inverse mix design of fly ash–GGBFS geopolymer concrete using a structured database of 672 specimens. Eight raw mix and curing variables were used after removing constant dosage-related columns. Thirteen regression models were first screened using 10-fold cross-validation, and the top models were further tuned using Bayesian optimisation. To move beyond point prediction, split conformal prediction, K-fold residual conformal prediction, and jackknife-after-bootstrap intervals were evaluated at 95%, 90%, and 80% confidence levels. GradientBoost achieved the highest cross-validated accuracy and was retained for residual diagnostics, Mondrian conformal analysis, convergent interpretability, and inverse design. LightGBM achieved the lowest held-out test error and, when paired with K-fold residual conformal prediction, provided the recommended 90% reliability-aware prediction configuration, with PICP = 0.9703, MPIW = 1.186&#xa0;MPa, and ES = 0.9739. Nearest-neighbour applicability-domain analysis showed that the high R² values mainly reflect dense interpolation within the structured FA–GGBFS experimental envelope rather than unrestricted extrapolation. SHAP, Morris, and PAWN analyses consistently identified curing time, recycled aggregate content, and aggregate composition as influential variables. Finally, local adaptive conformal thresholds were coupled with NSGA-II to identify six deduplicated strength–uncertainty design regions. The framework is useful for in-envelope reliability-aware screening of FA–GGBFS mixes, while transfer to new geopolymer chemistries requires external validation and recalibration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reliability-aware compressive-strength prediction and inverse mix design of fly ash–GGBFS geopolymer concrete via conformal prediction

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Gaurav Jagad,
  • Amit Thoriya

摘要

This study develops a reliability-aware machine-learning framework for compressive-strength prediction and inverse mix design of fly ash–GGBFS geopolymer concrete using a structured database of 672 specimens. Eight raw mix and curing variables were used after removing constant dosage-related columns. Thirteen regression models were first screened using 10-fold cross-validation, and the top models were further tuned using Bayesian optimisation. To move beyond point prediction, split conformal prediction, K-fold residual conformal prediction, and jackknife-after-bootstrap intervals were evaluated at 95%, 90%, and 80% confidence levels. GradientBoost achieved the highest cross-validated accuracy and was retained for residual diagnostics, Mondrian conformal analysis, convergent interpretability, and inverse design. LightGBM achieved the lowest held-out test error and, when paired with K-fold residual conformal prediction, provided the recommended 90% reliability-aware prediction configuration, with PICP = 0.9703, MPIW = 1.186 MPa, and ES = 0.9739. Nearest-neighbour applicability-domain analysis showed that the high R² values mainly reflect dense interpolation within the structured FA–GGBFS experimental envelope rather than unrestricted extrapolation. SHAP, Morris, and PAWN analyses consistently identified curing time, recycled aggregate content, and aggregate composition as influential variables. Finally, local adaptive conformal thresholds were coupled with NSGA-II to identify six deduplicated strength–uncertainty design regions. The framework is useful for in-envelope reliability-aware screening of FA–GGBFS mixes, while transfer to new geopolymer chemistries requires external validation and recalibration.