Two Scales, Hybrid Model for Soils, Involving Artificial Neural Network and Finite Element Procedure

Abstract

A hybrid ANN-FE solution is presented as a result of two level analysis of soils: a level of a laboratory sample and a level of engineering geotechnical problem. Engineering properties of soils (sands) are represented directly in the form of ANN (this is in contrast with our former paper where ANN approximated constitutive relationships). Initially the ANN is trained with Duncan formula (Duncan and Chang [2]), then it is re-trained (calibrated) with some available experimental data, specific for the soil considered. The obtained approximation of the constitutive parameters is used directly in finite element method at the level of a single element at the scale of the laboratory sample to check the correct representation of the laboratory test. Then, the finite element that was successfully tested at the level of laboratory sample is used at the macro level to solve engineering problems involving the soil for which it was calibrated.