Artificial neural network for predicting creep of concrete

Abstract

The concrete is today the building material by excellence. Drying accompanies the hardening of concrete and leads to significant dimensional changes that appear as cracks. These cracks influence the durability of the concrete works. Deforming a concrete element subjected to long-term loading is the sum of said instantaneous and delayed deformation due to creep deformation. Concrete creep is the continuous process of deformation of an element, which exerts a constant or variable load. It depends in particular on the characteristics of concrete, age during loading, the thickness of the element of the environmental humidity, and time. Creep is a complex phenomenon, recognized but poorly understood. It is related to the effects of migration of water into the pores and capillaries of the matrix and to a process of reorganization of the structure of hydrated binder crystals. Applying a nonparametric approach called artificial neural network (ANN) to effectively predict the dimensional changes due to creep drying is the subject of this research. Using this approach allows to develop models for predicting creep. These models use a multilayer backpropagation. They depend on a very large database of experimental results issued from the literature (RILEM Data Bank) and on appropriate choice of architectures and learning processes. These models take into account the different parameters of concrete preservation and making, which affect drying creep of concrete as relative humidity, cure period, water-to-cement ratio (W/C), volume-to-surface area ratio (V/S), and fine aggregate-to-total aggregate ratio, or fine aggregate-to-total aggregate ratio. To validate these models, they are compared with parametric models as B3, ACI 209, CEB, and GL2000. In these comparisons, it appears that ANN approach describes correctly the evolution with time of drying creep. A parametric study is also conducted to quantify the degree of influence of some of the different parameters used in the developed neural network model.