Neural network-based constitutive modeling of granular material

A new approach to material constitutive modeling using neural networks (NN) is applied to modeling the non-linear behavior of granular material. Neural networks develop material models through learning from examples and storing the underlying problem-dependent knowledge and information in its connection weights. This approach is used to model the drained and undrained behavior of sand using the results obtained from a series of uniform triaxial compression tests. The test results provide the stress-strain data required for the NN training.;The NN-based material modeling approach opens up new possibilities for developing material models from non-uniform material tests, where points in the specimen follow different stress paths. An autoprogressive algorithm is developed to extract the material constitutive behavior from the non-uniform material test. The autoprogressive algorithm uses an iterative process that consists of performing dual finite element analyses, one to apply one set of measured boundary conditions (i.e., boundary forces) and the other to enforce the second set of measured boundary conditions (i.e., boundary displacements), and a NN training. The simulation of the test using a finite element analysis, in which the material model is represented by the autoprogressively trained NN material model, matches both sets of the measured boundary conditions.;An autoprogressive training simulator is developed, and non-linear finite element analysis is implemented to handle geometrically non-linear problems. This simulator is then used for the autoprogressive training of the NN material models using the results of drained triaxial compression tests with end friction.