Direct Field Calibration for Model Simulations of Deep Excavations

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

Numerical modeling of geotechnical problems is used routinely in major construction projects. These models and simulations involve nonlinear analysis of staged construction for open-cut excavations, tunnels, slopes, and similar engineered structures. The most important and difficult part of these computer simulations is the representation of the constitutive behavior of the soil strata. In current engineering practice, the engineer selects an existing constitutive model and calibrates its parameters to match the results of few laboratory material tests. The tests do not generate information on important aspects of the soil behavior, which are relevant for a field problem. Often the results of numerical simulation using the calibrated constitutive models do not match the field measurements. Ad hoc methods are used to select and adjust the constitutive model and its properties to match field performance. We propose a novel, powerful and systematic method to calibrate the constitutive model of the soil behavior directly from field measurements. We will apply the autoprogressive method; a neural network based methodology that has been proposed by Ghaboussi and his co-workers, to the modeling of staged construction for a deep braced excavation. A neural network (NN) material model will represent the constitutive model of the soil behavior and will be calibrated using laboratory test and observed field behavior of excavations. Initially, the proposed methodology will be applied to synthetically generated "field measurements" from numerical simulations of deep excavations. The synthetic data will include wall lateral displacements and surface settlements. A classical bounding surface plasticity model will be used to represent clay behavior to generate the synthetic data. As a verification of the proposed approach, the soil behavior computed by the trained NN constitutive model can be compared to the classical soil model. The proposed methodology will then be applied to field measurements from deep excavations in Boston Central Artery/Tunnel CA/T project. The NN material will learn the constitutive model of the soil behavior directly from field measurements of deformations. The proposed approach can be applied to problems other than open-cut excavations. The approach will potentially greatly enhance the numerical modeling of geotechnical problems. Field observations and "local experience" can then be directly and systematically incorporated into numerical models.