ORALS

In geotechnics we rarely directly measure the parameters of the soils we need for engineering computations. Thus, the inverse problem that allows the needed values from directly measured data is frequently solved. Although the four presented strategies can be used to elaborate results of many geotechnical tests (such as CPTU or Marchetti dilatometer), in this summary we focus on one example only, namely the Falling Weight Deflectometer (FWD). This test is used to evaluate mechanical parameters of layered structures of road and pavements. Deflections due to an impulse load from the falling weight are measured in several points using geophones aligned on a rigid support. Determination of the mechanical parameters of the pavement layers is done by minimization of a mean square difference between the measured and a theoretical deflection. In the kernel of minimization procedures, there are costly computations of the theoretical deflections and their gradients (see [1]). The simplest use of Artificial Neural Network (ANN) in this context is the application of the well trained ANN as a surrogate of costly FEM computations in the minimization process. Once trained with limited number of results of the direct FE solutions the ANN gives the deflections for trial set of searched parameters. We use the ANN that approximates also the gradients with respect to its input. Such an ANN used it in frame of Truncated Newton method assures acceleration of the procedure. As a second strategy, we approximate directly the inverse relation between the set of parameters of the FE model of the layered structure and the deflection of its surface (see [2]). The input of the ANN is valued with the deflections and the output with the corresponding set of the model parameters. The ANN acts here as a universal approximate of an unknown functional relationship among the observed deflections and the searched parameters. Using various qualitative FE models, we can also train the ANN to discover qualitative properties of the structure which seems to be novel in the field of the inverse problem. Another possibility is the use of the ANN to directly approximate the inverse relation trained with the laboratory data collected from numerous real tests. We show that the necessary number of the laboratory test to train the ANN is reasonably small. The ANN trained with direct laboratory data acts here a special model-less form of phenomenological representation of constitutive relationships, based on observations (as in [3]).

References:
[1] P. Ruta, B. Krawczyk, A. Szydło, Identification of pavement elastic moduli by means of impact test, Engineering Structures, 100 (2015) 201-211
[2] M. Lefik and D.P. Boso, Inverse problem : soft solution, In Bytes and Science, Zavarise G. and Boso D.P. (eds.), CIMNE, Barcelona, Spain 2012.
[3] R.W. Meier. Backcalculation of Flexible Pavement Moduli from Falling Weight Deflectometer Data Using Artificial Neural Networks. Report prepared for U.S. Army, report GL-95-3, Washington DC, 1995.