The system based on the convolutional neural network IceNet predicts the state of the Arctic ice for months ahead. The tool will improve the early warning systems used to monitor the safety of animals and coastal settlements.
The Arctic is the area most susceptible to global warming. Predicting the dynamics of ice is a difficult physical task due to the fact that its formation is influenced by both the air above it and the ocean below it.
The model was trained on a dataset containing climate modeling data for several thousand years from decades of observations of climate change. IceNet is implemented in Python using TensorFlow. The IceNet input data is 50 monthly averaged climate variables. IceNet processes them using a series of convolutional blocks with batch normalization. The output data are ice density forecasts for the next 6 months.
The calculations were performed using the NVIDIA Quadro P4000 GPU, retraining the model on new data takes about a day. This is several thousand times faster than modeling based on the laws of physics. The accuracy of forecasts is 95% with a forecast period of two months in advance.