摘要 : With the development of experimental measurement technology and high-fidelity numerical simulations of granular materials, empirical-based classical constitutive models may not be able to take full advantage of the rapidly increas... 展开 With the development of experimental measurement technology and high-fidelity numerical simulations of granular materials, empirical-based classical constitutive models may not be able to take full advantage of the rapidly increasing available datasets. Machine learning-based models can inherently avoid phenomenological assumptions to directly learn the constitutive relationship from the datasets, and the trained model is sufficiently flexible to be reconstructed once new training samples are added. In this work, a coupled finite element method and machine learning (FEM-ML) computational framework is proposed for simulating granular materials. Gaussian process-based random loading paths and coupled FEM-DEM simulations are used to generate training samples. A parametrisation of the material deformation history is used to represent the historical influence of granular materials. An uncertainty-level-based active learning is utilised to evaluate the informativeness of data points for network training and then to establish an effective resampling scheme from a massive dataset. Two examples are provided to show the applicability of the implemented FEM-ML framework. The performance of the proposed framework is also evaluated, the error is systematically analysed, and possible improvements are discussed. The results demonstrate that the FEM-ML framework offers considerable improvements in terms of computational efficiency and competence to simulate the mechanical responses of granular materials. 收起