Parametric Assessment of Concrete Constituent Materials Using Machine Learning Techniques

Nowadays, technology has advanced, particularly in machine learning which is vital for minimizing the amount of human work required. Using machine learning approaches to estimate concrete properties has unquestionably triggered the interest of many researchers across the globe. Currently, an assessment method is widely adopted to calculate the impact of each input parameter on the output of a machine learning model. This paper evaluates the capability of various machine learning methodologies in conducting parametric assessments to understand the influence of each concrete constituent material on its compressive strength. It is accomplished by conducting a partial dependence analysis to quantify the effect of input features on the prediction results. As a part of the study, the effects of machine learning method selection for such analysis are also investigated by employing a concrete compressive strength algorithm developed using a decision tree, random forest, adaptive boosting, stochastic gradient boosting, and extreme gradient boosting. Additionally, the significance of the input features to the accuracy of the constructed estimation models is ranked through drop-out loss and MSE reduction. This investigation shows that the machine learning techniques could accurately predict the concrete's compressive strength with very high performance. Further, most analyzed algorithms yielded similar estimations regarding the strength of concrete constituent materials. In general, the study's results have shown that the drop-out loss and MSE reduction outputs were misleading, whereas the partial dependence plots provide a clear idea about the influence of the value of each feature on the prediction outcomes.