Comparative analysis of cement grade and cement strength as input features for machine learning-based concrete strength prediction
Jeonghyun Kim, Donwoo Lee, Andrzej Ubysz
Abstract
Machine learning (ML) has gained recognition as a valuable tool for predicting concrete properties. This study investigated the influence of input features related to cement strength on the performance of ML models. Four datasets with various features were prepared, and for each dataset, cement grade and cement strength were alternately applied as input features. ML models such as Random Forest, Extreme Gradient Boosting, and Multilayer Perceptron Neural Network were utilized to predict concrete strength for each dataset. The results showed a tendency for the prediction performance to improve when cement properties were used as input features, with the extent of improvement varying across the datasets. Permutation importance analysis indicated that cement strength often had a greater influence on the ML models than cement grade, positively enhancing prediction performance. Therefore, considering cement strength as an input feature is expected to be beneficial for constructing more accurate ML models.