RESEARCH ON HETEROGENEOUS SOLID STRESS MODEL BASED ON ARTIFICIAL NEURAL NETWORK

The Energy-Minimization Multi-Scale (EMMS) theory has been introduced into the multiphase particle-in-cell (MP-PIC) method to establish the heterogeneous EMMS solid stress model to account for the effect of non-uniform solid distribution. However, the calculation process is very complex and also very time consuming for this heterogeneous solid stress model. The expression of the heterogeneous EMMS solid stress can be obtained by manual fitting method. However, the fitting variable describes heterogeneous solid distribution as well as the fitting function describe the shape of solid stress are required for manually fitting. Since the heterogeneous solid stress function is highly nonlinear in nature, the fitting precision is not high enough for the manually fitting model. And there is an obvious deviation between the fitting correlation and the original EMMS solid stress, because it is hard to find out an appropriate parameter to characterize the heterogeneous solid concentration distribution as well as to find out an appropriate fitting function. In order to solve the above problems, an artificial neutral network (ANN) based machine learning method was proposed to avoid the characterization of the local distribution of solid volume fraction. Subsequently an ANN solid stress model which accounts for the detailed distribution of particle concentration was proposed to improve the fitting accuracy. Firstly, a two-marker based ANN solid stress model was established based on local particle concentration and particle non-uniform distribution index. Further, particle concentrations in the current cell and its neighboring cells were arrayed to represent the particle concentration distribution, thus to establish the ANN solid stress model based on particle concentration distribution. Then, the two models are compared with the EMMS solid stress model, and the effects of grid resolution and coarse-graining ratio on the model are also tested. The simulation results predicted with ANN model agreed well with that of the EMMS solid stress model, and the dependence of simulation results on grid resolution and coarse-graining ratio was also reduced.