Neural network modelling of properties of cement-based materials demystified

Publisher Website
Google Scholar
Abstract
Engineers often have to deal with materials of ill-defined behaviour such as cement-based materials in order to perform special design tasks. There is usually great difficulty in predicting the engineering properties of such materials due to various factors, including their non-homogeneous nature, their composite behaviour with dissimilar ingredients and sometimes the dual and/or contradictory effects of some components on the overall performance. Until recently, the methods used to predict the engineering properties of cement-based materials have been based mainly on statistical and mathematical models, which in turn are derived from human observation, empirical relationships and assumptions with limited ability to account for the effects of and interaction between all variables involved. An alternative approach, termed artificial neural networks (ANNs), has recently emerged in different engineering fields as a popular tool to predict the behaviour of materials. Due to the relatively recent adoption of ANNs for modelling the behaviour of cement-based materials, a good understanding of its fundamental basis and a critical assessment of its performance are essential. This paper examines the most widely used ANNs in materials modelling (the feed-forward, back-propagation (FFBP) neural networks). Guidelines for building, training, and validating such networks are provided. A critical assessment is presented of the effects of various parameters on the training and performance of FFBP networks and their use as an alternative approach to traditional modelling methods is evaluated through a case study. Recommendations are made to optimise the performance of ANNs. Engineers often have to deal with materials of ill-defined behaviour such as cement-based materials in order to perform special design tasks. There is usually great difficulty in predicting the engineering properties of such materials due to various factors, including their non-homogeneous nature, their composite behaviour with dissimilar ingredients and sometimes the dual and/or contradictory effects of some components on the overall performance. Until recently, the methods used to predict the engineering properties of cement-based materials have been based mainly on statistical and mathematical models, which in turn are derived from human observation, empirical relationships and assumptions with limited ability to account for the effects of and interaction between all variables involved. An alternative approach, termed artificial neural networks (ANNs), has recently emerged in different engineering fields as a popular tool to predict the behaviour of materials. Due to the relatively recent adoption of ANNs for modelling the behaviour of cement-based materials, a good understanding of its fundamental basis and a critical assessment of its performance are essential. This paper examines the most widely used ANNs in materials modelling (the feed-forward, back-propagation (FFBP) neural networks). Guidelines for building, training, and validating such networks are provided. A critical assessment is presented of the effects of various parameters on the training and performance of FFBP networks and their use as an alternative approach to traditional modelling methods is evaluated through a case study. Recommendations are made to optimise the performance of ANNs.