Back analysis of geomechanical parameters using classical and artificial intelligence techniques

Abstract

In this paper, a study is performed to evaluate the main advantages and limitations of different types of optimization algorithms for the identification of geomechanical parameters in underground structures. The main goal was to evaluate their main differences in terms of robustness and efficiency in diverse circumstances. Two different types of algorithms were tested: i) classical algorithms which use the gradient of the error function to guide the search, namely the steepest descent, conjugate gradient and quasi-Newton; and ii) an innovative evolutionary algorithm, from the artificial intelligence field, called evolution strategy. The first was coupled with a 3D model of a tunnel while for the latter analytical solutions were used for a much higher number of calculations and tests were expected to perform due to its nature. The algorithms were tested in both elasticity and elasto-plasticity using the Mohr-Coulomb constitutive model. In elasticity both types of algorithms showed a good performance but in elasto-plasticity the classical algorithms revealed a poor behaviour mainly in terms of robustness, failing to converge in several situations. It was found that in many situations these limitations were avoided by the innovative algorithm presented in this work showing an interesting potential for future developments.