AI4ER MRes Models & Forecasts

• Main Author: Andrew McDonald
• Format: Other/Unknown Material
• Language: English
• Published: Zenodo 2023
• Subjects: icenet; sea ice; deep learning; Sea ice
• Online Access: https://doi.org/10.5281/zenodo.8101236

A collection of model checkpoints and forecasts generated for the MRes thesis "Towards sharp sea ice concentration forecasts in the Arctic" by Andrew McDonald. See 4_forecast.ipynb and 5_evaluate.ipynb in the notebooks folder of the project GitHub repository at https://github.com/ampersandmcd/icenet-gan for a demo of how to make use of these files. Checkpoints radiant-sponge-59-great-unet-epoch=11-step=456contains the PyTorch model checkpoint of our best performing UNet model stilted-armadillo-99-great-gan-epoch=7-step=3024.ckptcontains the PyTorch model checkpoint of our best performing GAN model Forecasts radiant-sponge-59-great-unet.nc contains a forecast generated by our best performing UNet model for the test set beginning February 2018 and ending June 2019 stilted-armadillo-99-great-gan.nc contains a forecast generated by our best performing GAN model for the test set beginning February 2018 and ending June 2019 Why is the GAN forecast file so much larger than the UNet forecast file? The GAN forecast file is an ensemble forecast comprising 25 ensemblemembers, with a mean forecast cached as a 26th member. Hence, it is 26x the size of the deterministic UNet forecast. Completed in partial fulfilment of degree requirements for the MRes portion of the UKRI AI4ER CDT (https://ai4er-cdt.esc.cam.ac.uk/) based at the University of Cambridge (https://cam.ac.uk/).