Keywords: artificial neural network, multi-layer perceptron, approximation, parameter identification, affinity hydration model, cement paste.

A variety of engineering tasks leads to an inverse analysis problem. Generally, the aim of an inverse analysis is to rediscover unknown inputs from the known outputs. In common engineering applications, a goal is to determine the initial conditions and properties from physical experiments or, equivalently, to find a set of parameters for a numerical model describing properly the experiment.

While the numerical model of an experiment represents a well-defined mapping from input (model, material, structural, or other) parameters to output (structural response), there is no guarantee that the inverse relation even exist. In engineering practice the inverse relation is often ill-posed, highly nonlinear and multi-modal. Therefore, the choice of an appropriate identification strategy is not trivial. Moreover, such identification process is supposed to be performed repeatedly for any new measurement and therefore, the emphasis is also placed on the efficiency of the chosen identification method.

Overall, there are two main philosophies to the solution of identification problems [1]. A forward (classical) mode-direction is based on the definition of an error function of the difference between outputs of the model and experimental measurements. A solution comes with the minimisation of this error function. In the case of complex time-consuming models, the optimisation process can become easily unfeasible. Such a situation can be solved by introducing some approximation (surrogate or meta-) model.

The second philosophy, an inverse mode, assumes the existence of an inverse relationship between outputs and inputs. If such relationship is established, then the retrieval of the desired inputs is a matter of seconds and could be easily executed repeatedly. On the contrary, the main disadvantage is an exhausting search for the inverse relationship. Generally, this inverse model does not need to exist.

Nowadays, artificial neural networks (ANN) have became the most frequently used methods for the approximation of nonlinear relationships. In parameter identification problems they can be efficiently applied in both described identification modes. In the forward mode, ANN can be used to approximate the model reponse, while in the inverse mode they can approximate the inverse relation between model inputs and model outputs. In this paper, both applications are presented for parameter identification of the affinity hydration model for cement paste [2].

1

A. Kucerová, "Identification of nonlinear mechanical model parameters based on soft computing methods", Ecole Normale Supérieure de Cachan, Laboratoire de Mécanique et Technologie, Ph.D. thesis, 2007.

2

V. Šmilauer, "Multiscale Modeling of Hydrating Concrete", Saxe-Coburg Publications, UK, 2012. (To appear)

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £50 +P&P)