Deep-learning-based parametric calibration of discrete element method for the breakage simulation of sea ice

Abstract

The macro-scale material parameters of sea-ice and meso-scale model parameters in the discrete element method (DEM) for sea ice have a strongly nonlinear relationship because of the size effect in the DEM model. The parametric calibration is necessary to obtain high precision of sea-ice dynamics including the failure and fragmentation. This paper proposes a deep-learning-based parametric calibration for the parallel-bonding-based DEM model of sea ice, considering that the deep learning is good at establishing the nonlinear relationship of multiple input and output parameters. The training and prediction data are generated through DEM simulations, including uniaxial compression and three-point bending tests of sea ice in the DEM. The neural networks are employed to train the model by using the training data in which material parameters are the input data and model parameters are the output data. The prediction data illustrate that the prediction errors for different model parameters are less than 30%. The empirical formula that determines the bonding strength and internal friction from the compressive and flexural strength of sea ice is used for the validation as well. The comparison indicates that the neural networks have better precision than the empirical formula, and more parameters can be determined in the neural networks. Furthermore, the DEM simulation is used to validate whether the simulation results of strength can reach the input strength parameters. The validation shows that the error is lower than 6%. Hence, the proposed deep-learning-based parametric calibration yields highly accurate and effective results for DEM simulations.