Development of hybrid machine learning-based carbonation models with weighting function

Carbonation of concrete has significant influence on the service life of constructions and great effort has been made to establish an accurate and efficient model of carbonation that incorporates both internal and external factors. We present a hybrid machine learning (ML) approach that combined two single ML models: artificial neural network (ANN) and support vector machine (SVM). A literature survey generated a database containing 532 records of accelerated carbonation depth measurements for concrete mixtures inclusive of fly-ash blends. Six inputs comprising cement content, fly-ash replacement level, water-binder ratio (w/b), CO2 concentration, relative humidity and exposure time were selected for modeling, justified by grey relational analysis. The four ML models had excellent accuracy in predicting the carbonation depth of concrete, with the correlation coefficient ranging from 0.9788 to 0.9946, but the two hybrid ML models achieved superior performance to the single ANN and SVM models with characteristic higher correlation coefficients, lower mean value for the absolute error and lower standard deviation for its distribution. In addition, compared with other commonly known empirical carbonation models, the hybrid ML models showed more accurate carbonation depth prediction with smaller root mean square error. Furthermore, the weightings of the contributions of six selected factors to carbonation depth disclosed that CO2 concentration, w/b and binder content had higher relative importance to carbonation depth and should be given greater weighting in future carbonation model development.