Scholarlink

<p>Machine learning is commonly used to predict sorptivity from destructive test data, but its application in durability assessments of cover concrete using nondestructive test data remains limited due to initial surface moisture complexities. This study investigates seven ML algorithms (artificial neural network – ANN, Adaptive Neuro-fuzzy Interface System – ANFIS, random forest – RF, support vector regression – SVR, Gaussian process regression – GPR, eXtreme gradient boosting – XGBoost and decision tree – DT) to predict water sorptivity of cover concrete. 760 datasets of Surface Water Absorption Test (SWAT) results comprising OPC, GGBFS and fly ash concrete with 40%, 42%, 50% and 60% w/b contents were used. Seven parameters – unit coarse aggregate content (UAC), water-to-binder content (WBC), admixture dosage (Admix), cement type (CT), type of curing (TC), age at measurement (AM) and surface moisture content (SMC) – were input variables, while water sorptivity (WS) was the target output. DT model performed best in training, achieving a correlation coefficient (R) of 0.9354, mean absolute error (MAE) of 0.1082, mean squared error (MSE) of 0.0202 and variance accounted for (VAF) of 87.49%. SVR showed the weakest performance in training. In testing, ANN outperformed all other models with an R of 0.9189, MAE of 0.1305, MSE of 0.0298, and VAF of 84.44%, while ANFIS exhibited the poorest test performance. SHAP analysis identified UAC, TC, and SMC as the top features influencing model predictions. Using ANN, closed-form predictive equations were developed for practical application. The findings elucidate the advantages of non-invasive quality assessment, delivering practical insights for both material selection and maintenance strategies of concrete infrastructure.</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

Evaluation of machine learning models for predicting surface water absorption in cover concrete durability assessment

  • 期刊论文

  • 发表日期：

    2026年3月14日

查看全文
  • Raphael Nnodim Uwazuruonye,
  • Nafiu Olanrewaju Ogunsola,
  • Ajibola Ibrahim Quadri

摘要

Machine learning is commonly used to predict sorptivity from destructive test data, but its application in durability assessments of cover concrete using nondestructive test data remains limited due to initial surface moisture complexities. This study investigates seven ML algorithms (artificial neural network – ANN, Adaptive Neuro-fuzzy Interface System – ANFIS, random forest – RF, support vector regression – SVR, Gaussian process regression – GPR, eXtreme gradient boosting – XGBoost and decision tree – DT) to predict water sorptivity of cover concrete. 760 datasets of Surface Water Absorption Test (SWAT) results comprising OPC, GGBFS and fly ash concrete with 40%, 42%, 50% and 60% w/b contents were used. Seven parameters – unit coarse aggregate content (UAC), water-to-binder content (WBC), admixture dosage (Admix), cement type (CT), type of curing (TC), age at measurement (AM) and surface moisture content (SMC) – were input variables, while water sorptivity (WS) was the target output. DT model performed best in training, achieving a correlation coefficient (R) of 0.9354, mean absolute error (MAE) of 0.1082, mean squared error (MSE) of 0.0202 and variance accounted for (VAF) of 87.49%. SVR showed the weakest performance in training. In testing, ANN outperformed all other models with an R of 0.9189, MAE of 0.1305, MSE of 0.0298, and VAF of 84.44%, while ANFIS exhibited the poorest test performance. SHAP analysis identified UAC, TC, and SMC as the top features influencing model predictions. Using ANN, closed-form predictive equations were developed for practical application. The findings elucidate the advantages of non-invasive quality assessment, delivering practical insights for both material selection and maintenance strategies of concrete infrastructure.