A Study Of The Neural Network Model For Constitutive Relations For Granular Soils

More than 100 constitutive models of soil have already developed. Although the mechanical properties of soils have already been know clearer and the commonly using constitutive models are basic to represent the stress-strain characteristics of soils, it is also necessary to investigate new modeling method. The neural network method is the important research result of the artificial intelligence, has very strong nonlinear map ability,the ability of neglecting redundancy errors and fuzzy computing, is just because of such characteristics, make researching constitutive relations of soils in neural network method possible.Granular soils,because of its good engineering characteristics, are applied extensively in the railroad foundation,the side slope,the artificial island etc.The constitutive relations of granular soils are very complicated,under high confining pressures,granular soils are characterized by strain hardening and volumetric compression,while under low confining pressures it exhibit strain softening, volumetric dilatancy and nonlinear respectively. None of constitutive models can respond all characteristics of granular soils by now. The process of simulating constitutive relations of granular soils with neural network method and some problems appearing in the process was discussed in the first chapter, such as convergence of estimating curve, network-model's extending ability etc. The second chapter in first introduces the concept, the calculation ability of the neural network method in detail, and some applications in the soil engineering. The third chapter studys how to use the neural network method simulating constitutive relations, the problems include choice of the network-model's structure, choice of the network-model's training way and parameter, choice of the network-model's input and output parameters. The neural network model learned five sets of triaxial experiment data, and the result of learning indicated that the neural network model with reasonable structure and training rule is good to imitate constitutive relations of granular soils.The network model overcome the difficulty of measuring soil parameter, and also consider the influence of different stress-strain path, to some extent, the network model is better than traditional modeling method.