Prediction of the shear strength of lightweight concrete beams without web reinforcement based on a machine learning model optimized by a genetic algorithm

Abstract

The shear performance of concrete beams is affected by several factors, and some of effective variables and control mechanisms should be considered in predicting shear strength. Compared with ordinary concrete, lightweight concrete (LWC) is more unique, which greatly increases the difficulty of prediction. This is also why it is necessary to introduce machine learning (ML) technology with stronger data processing capabilities to predict the shear performance of LWC. Therefore, the aim of this study is to employ four well-known machine learning methods to predict the shear performance of LWC beams, namely, support vector machine (SVM), artificial neural network (ANN), radial basis function (RBF) neural network and random forest (RF). First, the shear test results of 220 LWC beams were collected, and the corresponding datasets were established based on possible influencing factors. Then, the optimal dataset partitioning methods for four different ML models were explored and determined. The parameter search process of the four models under the optimal dataset partitioning ratio was optimized using a genetic algorithm (GA), and four GA-ML prediction models were developed. The results show that the developed GA-RF model achieves a train set accuracy of 95.3 % and a test set accuracy of 93.3 %. The GA-ML technology significantly improves the accuracy of the prediction of shear strength compared to the formulaic model based on current international design standards. The prediction mechanism of the model is further analyzed to ensure that the model is reliable. The results indicate that it can be applied to predict the shear strength of LWC. In summary, the GA-RF model and importance analysis proposed in this article are practical methods for predicting and explaining the shear strength of LWC.