Microplane model parameters estimation using neural networks

Concrete is one of the most frequently used material in Civil Engineering. Nevertheless, as a highly heterogeneous material, it shows very complex non-linear behavior, which is extremely difficult to describe by a sound constitutive law. As a consequence, a numerical simulation of response of complex concrete structures still remains a very challenging and demanding topic.

One of the most promising approaches to modelling of concrete behavior is based on the microplane paradigm [

1

]. It is a fully three-dimensional material law that incorporates tensional and compressive softening, damage of the material, supports different combinations of loading, unloading and cyclic loading along with the development of damage-induced anisotropy of the material. As a result, the material model [

1

] is fully capable of predicting behavior of real-world concrete structures, once provided with proper input data. The major disadvantages of this model are, however, a large number of phenomenological material parameters and a high computational cost associated with structural analysis even in a parallel implementation [

2

].

The authors of the microplane model proposed a heuristic calibration procedure [

1

], that is based on the trial-and-error method but is computationally inefficient. Therefore, a new procedure based on artificial neural networks is proposed in the present contribution.

In order to asses the reliability of identified material parameters, results of a stochastic sensitivity study based on the Latin Hypercube Sampling (LHS) method are presented first. Different tests, proposed in [

2

], are simulated numerically and used to determine, which model parameters can be reliably identified from these tests. In the next step, a neural network-based procedure is presented for identification of material parameters. A crucial point is the generation of a training set used to determine weights of individual neurons. To this end, the LHS method is again employed as it allows using a limited number of computational simulations while ensuring the representativeness of the generated training set. The training procedure itself is based on a real-coded genetic algorithm SADE [

3

]. Finally, the application of the proposed identification procedure to the back analysis of laboratory experiments is presented.