Machine learning techniques versus classical statistics in strength predictions of eco-friendly masonry units

Abstract

Earth-based materials demonstrated promising characteristics in the development of eco-friendly, low cost and sustainable construction materials. However, their unconventional utilization in construction makes the assessment of their properties very difficult and inaccurate because they are assessed based on conventional materials procedures. Hence, the properties of earth-based materials are not well understood. The assessment of earth-based materials properties for sustainable construction is time-consuming, expensive, and inaccurate. To obtain more accurate properties, an artificial neural network and statistical linear regression analysis were used to predict the compressive strength of alkali-activated soil. Statistical linear regression analysis was carried out to compare the efficiency of the machine learning technique with the classical statistics model. Parameters such as Si/Al, activator level, curing temperature, water absorption, and weight were used as input parameters to predict the target variable. The coefficient of determination was used to examine the performance of the models. The results depict that artificial neural network outperformed statistical linear regression analysis with R2=0.74, RMSE=0.119 and R2 =0.48, RMSE=0.466 respectively. This indicates that statistical linear regression analysis is inefficient for prediction of the strength in alkali activated soils