In the Arctic, sea ice is undergoing for several decades a dramatic decline both in terms of spatial extent and average thickness. As sea ice strongly affects the exchanges of energy and momentum between the ocean and the atmosphere, its decline is accompanied by a particularly strong warming in polar regions, particularly in the Arctic. We will show that mechanical processes likely play a key role in this recent evolution.
Then, in order to understand and forecast the sea ice retroactions on climate evolution, we need to simulate its dynamic. We present here the modeling process and discretization strategy developed to build a simulation tool: FloeDyn. The goal was to build a tool enable to simulate sea ice at elementary scale (floes) and be able to link these simulations with large scale, it means to be able to perform simulation with a very large number of floes in interaction and able to be destroyed by chocks, waves or thermodynamic processes.
In the framework of the simulation of destruction process of floes via percution, we propose a solution to tackle a computational problem: detection of floe destruction via the Griffith energy minimization would totally destroy the efficiency of the algorithm implemented in FloeDyn. This solution, based upon development of neural networks, will avoid a large number of complex computations and preserve the performances of the computation code. In this exposé we will give a presentation of the ideas to build the neural network and optimize learning.