Predictive models for the added mass of model ice floes using supervised learning

Le résumé fourni par la source

Hydrodynamic effects need to be considered to achieve high fidelity in kinematic simulations of ice floes. Interactions with the surrounding fluid give rise to added mass, which can be modelled by coefficients proportional to the acceleration of the moving ice floe. These coefficients are usually obtained from simple heuristics, model testing or via computational fluid dynamics. However, heuristics are usually restricted to simple geometries and based on significant simplifications, while model testing and computational fluid dynamics are both complex and time-consuming, making all approaches impractical for the accurate and efficient simulation of ice fields with a large number of floes having diverse geometries. Here, we create a dataset of added mass coefficients of prismatic model ice floes floating horizontally at still water level using a boundary element method and introduce two machine learning models for the prediction of the added mass tensors of floes from their geometries. By analyzing general trends in the dataset, we find that changes to the floe geometry systematically affect the tensors and that different tensor entries dominate during undisturbed motions and during collisions. The introduced machine learning models are capable of predicting any individual entry of the learned tensors with less than 1% error, with the convolutional neural network being able to predict added mass for floes of all tested geometries. Our approach enables the kinematic simulation of large and complex model ice fields.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.