Compressive Strength Prediction of Alkali‐Activated Slag Concretes by Using Artificial Neural Network (ANN) and Alternating Conditional Expectation (ACE)

Abstract

Compressive strength of alkali‐activated slag (AAS) concrete is influenced by multi‐factors in a nonlinear way. Both artificial neural network (ANN) and alternating conditional expectation (ACE) models of 3‐day (3 d) and 28‐day (28 d) compressive strength of AAS were established in this study by using the data reported in related literature, where alkali concentration of activator (Na2O%), modulus of activator (Ms), water/binder ratio (W/B), surface area of slag (SA), and basicity index of slag (Kb) were taken as input parameters. The models were employed later to predict 3 d and 28 d compressive strength of AAS concretes, respectively, and the results were validated by experimental work. The results show that both the ANN and the ACE models had adequate accuracy, no matter 3 d or 28 d compressive strength was considered. Compared to the 3 d compressive strength, due to data scattering that increased with the increase of data size, both the models did not yield a higher accuracy in the case of 28 d strength. However, also due to the increase in data size, both the models were more feasible to implement 28 d strength prediction as a result of sufficient learning and training during modeling. In addition, based on ACE analysis, the weight‐influencing compressive strength of AAS decreased in a sequence of Na2O% > Ms > W/B > Kb > SA. If data size was sufficiently large, it was more suitable to establish an ANN model for compressive strength prediction of AAS concretes. Otherwise, ACE could be considered as an alternative to yield an acceptable result.