An Artificial neural network for predicting accelerated and natural carbonation depths in blast-furnace slag and fly-ash based concretes

Abstract

Carbonation is one of the main depassivation phenomena in reinforced concrete. The prediction of this phenomenon presents a major importance, especially in the case of concrete containing low carbon binders. This paper develops a methodology using an artificial neural network (ANN) to compute the carbonation depth for different concretes. Twenty-six sources were used to build the database, allowing the obtention of 2537 sets of data. Besides, one of the main advantages of the present method is to enable the consideration of accelerated and natural test conditions with a unique tool. This way, equivalences between accelerated and natural carbonation rates can be established. Twenty variables are considered as input parameters, concerning essentially the composition of the concrete material (with clinker, filler, slag, fly ash, sand, aggregate and water content as well as sand and aggregate quality), the environmental conditions of the curing treatment, the preconditioning and the carbon dioxide exposure (time, temperature, relative humidity and partial pressure of CO₂). The learning and validation of the results allowed the obtention of determination coefficients of 0.98 and 0.93 respectively. The purpose of this tool being its application in an operational context, verifications on highroad structures are performed and lead to satisfactory results (mean relative errors inferior to 15%).