Intelligent identification of the Lattice Discrete Particle Model (LDPM) parameters using neural networks

Abstract

One of the primary concerns in structural engineering is the prediction of response of concrete structures under various environmental and loading conditions. The accurate prediction of the mechanical behavior of concrete requires a comprehensive understanding of the mechanical properties of concrete and the factors that influence these properties. The need to predict the behavior of concrete, to solve practical engineering problems, has attracted researchers to develop constitutive models, which attempt to replicate the behavior of concrete. The complex nature of concrete necessitates the use of comprehensive models that accurately describe the mechanical behavior of the material. The model used in this study is the Lattice Discrete Particle Model (LDPM) (Cusatis et. al). LDPM is a comprehensive meso-scale model that accounts for the heterogeneity of concrete and its influence on the evolution of damage and fracture. LDPM has shown exemplary and unprecedented performance in modeling and predicting concrete behavior subjected to various types of loading.

Such comprehensive models are governed by a number of model parameters. The available experimental data does not provide complete information to calibrate these models, which leads to an inaccurate estimation of the parameters. An identification system that uses the limited available data to identify the governing model parameters is thus necessary to calibrate such comprehensive models. The use of artificial intelligence as a tool for parameter identification has gained popularity over the last few decades. Artificial Neural Network (ANN) has been used in the past to predict fracture parameters. However, ANN has not been used as a parameter identification tool for comprehensive models such as the LDPM. The application of ANN for approximating functions between input and desired output variables has been advantageously used in this research to identify the LDPM parameters. A multi-step procedure is proposed to identify the LDPM parameters, with limited information about the concrete properties used in the experiment, using a feed-forward backpropagation neural network. The missing concrete properties are estimated based on ACI Standards and relevant experimental literature.