Prediction of flexural and uniaxial compressive strengths of sea ice with optimized recurrent neural network

Abstract

The mechanical properties of sea ice are important parameters for determining the ice load on engineering structures, such as offshore platforms, ships and ports. In the investigations of the mechanical properties of sea ice, experimental data samples are few, whereas the results predicted from an empirical formula feature a significant error and thus do not satisfy engineering requirements. Therefore, in this study, the stochastic gradient descent method combined with the Adam optimization algorithm is used to optimize a recurrent neural network (RNN), and an RNN model is established to predict the bending and compressive strengths of sea ice. This model can output the mechanical properties of sea ice using only physical parameters such as sea ice temperature, salinity, density, and loading rate. The overall average error in the prediction of the bending and compressive strengths of sea ice is less than 20%. The model is compared with an empirical formula, back-propagation neural network, and genetic-algorithm back-propagation neural network to predict the mechanical properties of sea ice. The results show that the model is optimized in terms of goodness-of-fit, mean relative error, mean absolute error, and mean square error, which implies that the optimized RNN is more suitable for predicting the bending and compressive strengths of sea ice. The model can be applied to rapidly and accurately predict the mechanical properties of sea ice using easily measured physical sea ice parameters, thus providing a reference for the anti-ice design of ocean engineering structures.