The prediction of durability to freeze-thaw of limestone aggregates using machine-learning techniques


CONSTRUCTION AND BUILDING MATERIALS, vol.324, 2022 (Journal Indexed in SCI) identifier identifier

  • Publication Type: Article / Article
  • Volume: 324
  • Publication Date: 2022
  • Doi Number: 10.1016/j.conbuildmat.2022.126678
  • Keywords: Limestone aggregate, Freeze-thaw resistance, Gaussian process regression, Support vector machine, Regression trees, SUPPORT VECTOR REGRESSION, COMPRESSIVE STRENGTH, RECYCLED CONCRETE, RANDOM FOREST, MODEL, ROCK, DETERIORATION, TREE, CLASSIFICATION, RESISTANCE


To design structures such as buildings, roads, highways, bridges, and railroads, one of the most important parameters is the freeze-thaw resistance of aggregate. The direct analysis of freeze-thaw resistance in the laboratory causes losses of both electrical energy and time. Also, performing this analysis by manual methods may lead to experimental errors. To overcome these problems, it has been proposed to predict freeze-thaw resistance of aggregates with machine learning techniques. Different aggregate samples were collected from the Cukurova region in Turkey, and uniaxial compressive strength (UCS), specific gravity (SG), water absorption rate (WA), Los Angeles abrasion (LA), P-wave velocity (Vp) and freeze-thaw resistance (F-T) tests were applied to these samples. The results obtained were recorded and the data set containing 220 data was created. In this study, four different machine-learning techniques such as Gaussian Process Regression (GPR-Exponential and GPR-Matern5/2), Support Vector Machine (SVM) and Regression Trees (RT) were performed to freeze-thaw prediction of aggregate. Coefficient of determination (R2), root mean squared error (RMSE), mean square error (MSE), and mean absolute error (MAE) were selected as performance assessment metrics for prediction. The results demonstrate the successful application of machine-learning techniques in predicting freeze-thaw resistance. GPR-Exponential is the best prediction model that has the highest R2 (0.989, 0.958), lowest RMSE (0.091, 0.067), MSE (0.008, 0.004) and MAE (0.052, 0.046) in both training and testing stages, respectively.