Alkali-silica reaction (ASR) and expansion induced are one of the major concerns over durability of concrete structures. It causes an increase in internal stresses and hence deteriorating mechanical properties, reducing load capacity, and compromising structures’ serviceability. Factors governing ASR include water-to-cement ratio, alkali content, the content of reactive aggregate, exposure temperature, relative humidity, etc. Previous prediction models for ASR expansion focus on the measurements under particular conditions and these models are suffering from the neglect of effects of some factors. This study aims to propose a back-propagation neural network (BPNN) model to predict ASR by considering both mix compositions and environmental conditions. In this paper, by adopting two different training functions, namely Levenberg-Marquardt function (LM) and Bayesian regularisation backpropagation (BR), two models of LM-BPNN and BR-BPNN were developed and proposed. To achieve this, a database containing 168 pairs of data from 25 studies was established, its data includes water-to-cement ratio, cement content, alkali fraction, content of non-reactive and reactive fine/coarse aggregate, temperature, and relative humidity. Results indicate that both LM-BPNN and BR-BPNN models can produce reasonable estimates for ASR expansion at 180 days with MSE and R of 0.00209%, 0.002753%, 0.926 and 0.912, respectively. Results reveal that fraction of reactive fine/coarse aggregate is dominant on ASR expansion while alkali content and humidity have a significant correlation with expansion. Water-to-cement ratio, cement content, and content of non-reactive fine/coarse aggregate have low impact on ASR expansion based on the results.
Read full abstract