Comparing durability and compressive strength predictions of hyperoptimized random forests and artificial neural networks on a small dataset of concrete containing nano SiO2 and RHA

Abstract

Through the increasing use of supplementary cementitious materials, the properties of concrete have taken on increased significance in a design code. Using reliable prediction models based on a small data set for the mechanical properties and durability of concrete can reduce the number of trial batches and experiments needed to produce useful design data in the laboratory, reducing time as well as resources. In this study, we investigate how the properties of water penetration, chlorine resistance, and compressive strength can be predicted by polynomial regression (PR), random forest (RF) regression, and artificial neural networks (ANNs) based on the input values of density, workability, and the constituent amount of rice husk ash, cement, and nano SiO 2 . We vary the training data used and test the coefficient of determination ( R 2 score) on the remaining data as a test set to measure predictive capability. We show that RFs and ANNs outperform PR in all settings and have unambiguously extrapolating properties when hyperparameter optimization is designed for this purpose. Remarkably, we obtain R 2 scores on the test data of 0.858 − 0.990 for RFs and 0.825 − 0.985 for ANNs.