Prediction of ice resistance for ice-going ships in level ice using artificial neural network technique

Ice resistance is affected by many parameters, i.e., ship geometries, ice properties and interaction properties. Over the years, various methods ranging from empirical, semi-empirical, numerical methods and various of their combinations were developed to calculate the ice resistance with certain degree of success. In this study, from a brand-new perspective, a data-driven approach for estimating ice resistance based on the artificial neural network was proposed. The artificial neural network (ANN) is one of the main tools for machine learning. It can accurately and efficiently correlate directly the inputs and outputs of complex and nonlinear systems, such as the ice resistance calculation system. This physical system involves several interaction phases and at least three groups of input parameters (i.e., ship, ice and interaction characteristics). Following the basic idea of ANN, we trained and built six ANN models based on datasets that were collected over past decades involving both model tests and full-scale measurements. The six different ANN models differ in the amount of input parameters. Based on comparative studies, among all the input parameters (i.e., 7 variables in total: ship length, ship breadth, ship draft, stem angle, ship speed, ice flexural strength and ice thickness), we found that the ship breadth, ice thickness and ship speed have the largest influence on the calculation of ice resistance. Afterwards, the 7-variable ANN model's prediction was compared with existing semi-empirical methods and measurements; and favorable agreement was achieved with fairly simple matrix form formulas (i.e., Eq. (19)). The formula offers another simple, yet reliable approach to calculate ice resistance. However, since the method is data driven, high quality data are always needed in improving the predicting capability of the relevant ANN model following the methodology outlined in this paper.