Enhancing the reliability and accuracy of machine learning models for predicting carbonation progress in fly ash‐concrete: A multifaceted approach

Abstract

AbstractExisting machine learning models for carbonation in fly ash blended concrete, developed and validated using accelerated or combined datasets, lack validation against natural carbonation processes in concrete. Furthermore, the reliability and accuracy of these models are directly influenced by the sets of input variables used in model training. This research specifically aims to investigate the reliability and accuracy of machine learning models, trained with accelerated datasets, in predicting the natural carbonation process. Additionally, this study introduces Gaussian process regression (GPR) as an alternative modeling tool, due to its potential for enhanced prediction accuracy. The results of this study indicate that GPR can satisfactorily predict natural carbonation progress in fly ash concrete. Moreover, the study carried out a model interpretation to provide intuitive feature contributions in the black‐box model. The results indicate that the accuracy of the natural carbonation process evaluation of models trained with accelerated carbonation dataset is less than 0.2 R2 value, indicating the inappropriateness of using accelerated carbonation data to train reliable models. Additionally, the model interpretation shows a significant dependency of the natural carbonation process on the binder's chemical composition, highlighting key chemical factors influencing carbonation.