| Constitutive models of geomaterials are vital elements in computational soil mechanics for solving boundary value problems (BVPs). Classical constitutive models based on elastoplasticity theories are constrained since their development is influenced by compromise between complexity, accuracy, practicality, and implementability in numerical solutions. A successful solution of BVP is achieved by realization of accurate constitutive model, efficient numerical solution technique, and availability of computational resources. Unfortunately, regardless of the degree of sophistication of numerical solutions or digital computing resources used, less accurate constitutive models will lead to considerably erroneous solutions.; This research applies an alternative approach that utilizes adaptive learning of neural networks to material's mechanical behavior. The constitutive models are developed by training on experimental data obtained from ordinary tests at desired conditions. Unlike classical approaches that require specialized tests for model calibration, and apply ill-founded assumptions and drastic approximation, well trained models can reflect precisely the internal physical processes governing behavior of material. It is this feature that makes neural network-based material modeling superior to classical approaches.; To verify this hypothesis, a two-stage modeling program was performed. First, several modeling methodologies were proposed to approximate response of theoretical functions of variable complexity. Because data generated from theoretical functions are noise-free, any prediction inaccuracy will be exclusively attributed to the used methodology. In second stage, best methodology (true sequential dynamic mapping) obtained from previous stage was applied to model constitutive behavior of fine-grained soils in monotonic and cyclic loading. Fourteen fine-grained soils differing in their geotechnical properties were obtained from several locations in Kansas. A total of 124 strain-controlled monotonic and cyclic uniaxial compression tests on these soils were preformed. The neural network-based constitutive models developed for monotonic loading included effect of compaction state and soil properties on mechanical behavior. The cyclic models for one soil included effect of number of hysteretic cycles and compaction state. The developed models indicated high accuracy in predicting soil behavior including softening as compared to experimental stress-strain responses.; Neurosimulators were also developed in which the neural network material models comprise one component. The neurosimulators operate by starting from given initial stress-strain state when provided with known basic information about the system such as compaction condition, soil properties, locations, extent, and number of unloadings, and simulation resolution. The developed cyclic and monotonic neurosimulators were very efficient in simulating real behavior of soils as judged (quantitatively/qualitatively) by both experimental observations and known logical trends.; Methods for enhancing soil modeling methodologies such as model inversion, model expansion to 3-D and for accounting for stress-path dependency, and implementation in numerical solution techniques were also proposed. |
