Prediction of alkali-silica reaction expansion of concrete using artificial neural networks

Abstract

This paper presents a hybrid machine learning method for the prediction of concrete expansion induced by alkali-silica reaction (ASR) and assembles a comprehensive and reliable experimental database comprising of around 1900 sets of ASR expansion data from literature to calibrate and validate the machine learning-based prediction model. The hybrid machine learning method employs a beta differential evolution-improve particle swarm optimization algorithm (BDE-IPSO) to tune weights and biases of the artificial neural network (ANN) model. The model adopts 11 variables as input, in terms of material composition, specimen geometry and environmental conditions, and can predict ASR expansion with great accuracy. The results demonstrate that the established prediction model is able to capture all available experimental aspects of ASR expansion, including: (a) effects of reactivity, size, content of reactive aggregate, water-to-cement ratio, and alkali concentration; (b) effects of temperature and relative humidity; (c) size effects of specimen geometry; and (d) the time-dependent behavior.