Comparative use of different AI methods for the prediction of concrete compressive strength

Abstract

Concrete mix design requires specialized knowledge and techniques for characterization. However, this process is time-consuming, and the mechanical properties, such as strength, can vary due to factors like cement type, water content, aggregates, and curing time. Additionally, analytical mathematical models are often used to estimate concrete characteristics. However, accurately determining concrete properties without laboratory testing is challenging, especially when nontraditional materials, such as certain supplementary cementitious materials, are involved. Recently, artificial intelligence has become a powerful resource that enables machine learning-based forecasting using available data. This study utilized RapidMiner® software to design models capable of analyzing various types of tagged data and performing machine learning predictions. These models were applied to over 5,373 concrete formulations compiled from 137 literature sources. The simulations used artificial neural networks or deep learning, generalized linear, decision tree, random forest, support vector machine, and gradient-boosted tree models to predict the compressive strength of 8 concrete mix designs containing different SCMs. The accuracy of models was estimated using traditional statistical indices such as R², MAPE and RMSE. The most accurate model was found to be a gradient-boosted tree followed by deep learning and random forest. Forecasts were validated with high accuracy by comparing experimental results to numerical data.