Prediction of Concrete Modulus of Elasticity Using Deep Learning

Abstract

Modulus of Elasticity (Ec’) is a key parameter in structural engineering concrete designs. In concrete as a composite material, Ec’ is a function of compressive strength and the proportions of components in the concrete matrix (percentages of aggregates and cement). The inaccuracy and dispersity in estimating Ec’ from models provided by the existing codes of practice strongly affect the performance and design of the concrete structures. In this study, a dataset of 189 experimental concrete compressive strength results were collected from the available literature. The data set includes curing time (in days) for the concrete specimens, concrete density, experimental compressive strength (fc’), experimental Ec’ and several additives (e.g., slag, gypsum…etc.) with a total of 13 variables. Deep artificial neural networks (DANN) were used to model and analyze the effects of these variables on Ec’. A grid search over 2 hidden layers of DANNs was conducted to compute the best performed DANN. A total of 49 DANN models were developed in this study to predict concrete Ec’. The best performed DANN had a coefficient of determination (R2) of 0.81 and was selected for further analysis. Importance scoring was performed on the best DANN and results revealed that compressive strength had the highest importance score followed by water/cement ratio (w/c). Interestingly, the specimen sizes and curing days had the 6th and 8th scoring respectively from the 13 investigated variables. Ground pumice had the highest scoring compared to other additives. Sensitivity analyses were conducted revealing that at low specimen sizes of 10 mm, the Ec’ may vary by ~50%, while at higher size (150 mm), the Ec’ had less scatter and more reliable values.