ANN-based predictive mimicker for the constitutive model of engineered cementitious composites (ECC)

Abstract

The enhanced mechanical and durability properties of Engineered Cementitious composite (ECC) make it famous worldwide. However, due to the unavailability of material design guides, its mixed design is still based on extensive experimentation, which is a highly uneconomical and time-consuming process. This study aims to develop a machine learning-based model capable of predicting the suitable mix design for ECC. Initially, a dataset of 147 data points composed of mixed design and their associated stress-strain curves was collected from the published literature. This paper uses an Artificial Neural Network (ANN) model incorporating 10 input parameters, including all the ingredients of the mix design and the properties of the fibers utilized to predict the complete tensile stress-strain curve of ECC in terms of the bilinear idealized curve. Furthermore, a detailed sensitivity analysis was also conducted to study the influence of each parameter on the tensile behavior of ECC. This model gives RMSE for a training set of yield stress, yield strain, ultimate stress, and ultimate strain as 0.042, 0.002, 0.36, and 0.3, respectively, while for the testing set, RMSE is 0.2, 0.004, 0.374, 0.3133 and 0.18, respectively. Finally, the model's performance was revalidated experimentally by employing mixes that were not part of previous data. The results show that this model possesses a high accuracy and can help to design ECC without extensive experimentation, which can help in the advancement of the commercialization potential of this robust composite.