Scholarlink

<p>Bituminous binders for building envelopes must resist flow at elevated temperature while preserving extensibility for crack-bridging under service conditions. This work establishes and validates a predictive framework for ductility at 25&#xa0;°C of LDPE modified VG-30 bitumen from four routine inputs—LDPE content (wt %), penetration at 25&#xa0;°C (dmm), softening point (<sup>o</sup>C), and specific gravity—using a uniformly generated laboratory corpus of 1,100 observations (training = 770; testing = 330). The correlation map reveals a structured dependency in which ductility aligns strongly and positively with penetration, and negatively with LDPE content (wt %) and softening point, motivating nonlinear modeling. Five machine learning based modelling techniques—Random Tree (RT), Reduced-Error Pruning Tree (REP Tree), Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Machine (SVM)—were benchmarked under a cross-validated with a held-out test set method. A leakage audit confirmed that neither cross-validation folds nor the hold-out set shared engineered/augmented rows or replicate sub-specimens from the same batch, ensuring zero train–test contamination. SVM_RBF achieved the most balanced fidelity and generalization (training: CC = 0.9887, MAE = 0.3241, RMSE = 0.4956, SI = 0.0187, NSE = 0.9778; testing: CC = 0.9877, MAE = 0.3933, RMSE = 0.5282, SI = 0.0203, NSE = 0.9749), with agreement plots, box plots, relative-error envelopes, and a Taylor diagram confirming close alignment with measurements. A leave-one-input-out sensitivity analysis on the testing set ranked penetration at 25&#xa0;°C (dmm) as the dominant driver, followed by LDPE content (wt%), while softening point and specific gravity added limited incremental value once the former were included. The framework provides a compact and explainable route for screening LDPE content (wt%)-modified binders for roofing and waterproofing: set target ductility using SVM_RBF predictions, verify interaction patterns with the heatmap, and control temperature susceptibility (penetration at 25&#xa0;°C) and LDPE content (wt%) as the primary levers for mix design and quality control.</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

Ductility prediction for LDPE-modified VG-30 bitumen in building-envelope applications using RT, REP Tree, RF, ANN, and SVM

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Waseem Akram,
  • Pushpendra Kumar Sharma,
  • Bishnu Kant Shukla

摘要

Bituminous binders for building envelopes must resist flow at elevated temperature while preserving extensibility for crack-bridging under service conditions. This work establishes and validates a predictive framework for ductility at 25 °C of LDPE modified VG-30 bitumen from four routine inputs—LDPE content (wt %), penetration at 25 °C (dmm), softening point (oC), and specific gravity—using a uniformly generated laboratory corpus of 1,100 observations (training = 770; testing = 330). The correlation map reveals a structured dependency in which ductility aligns strongly and positively with penetration, and negatively with LDPE content (wt %) and softening point, motivating nonlinear modeling. Five machine learning based modelling techniques—Random Tree (RT), Reduced-Error Pruning Tree (REP Tree), Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Machine (SVM)—were benchmarked under a cross-validated with a held-out test set method. A leakage audit confirmed that neither cross-validation folds nor the hold-out set shared engineered/augmented rows or replicate sub-specimens from the same batch, ensuring zero train–test contamination. SVM_RBF achieved the most balanced fidelity and generalization (training: CC = 0.9887, MAE = 0.3241, RMSE = 0.4956, SI = 0.0187, NSE = 0.9778; testing: CC = 0.9877, MAE = 0.3933, RMSE = 0.5282, SI = 0.0203, NSE = 0.9749), with agreement plots, box plots, relative-error envelopes, and a Taylor diagram confirming close alignment with measurements. A leave-one-input-out sensitivity analysis on the testing set ranked penetration at 25 °C (dmm) as the dominant driver, followed by LDPE content (wt%), while softening point and specific gravity added limited incremental value once the former were included. The framework provides a compact and explainable route for screening LDPE content (wt%)-modified binders for roofing and waterproofing: set target ductility using SVM_RBF predictions, verify interaction patterns with the heatmap, and control temperature susceptibility (penetration at 25 °C) and LDPE content (wt%) as the primary levers for mix design and quality control.