The module adddresses oceanic mesoscale eddies analysis by providing the tools for pre-processing of FESOM2 data and training GNN models.
The development of the Python packages is currently underway. In this version a pre-processing pipeline has been developed, with both a complete notebook and a demonstration one. They allow interpolations from unstructured to regular grids and vice versa to prepare the dataset for the training. The actual training has both a notebook for demonstration with plots and a Python script to be run with Slurm on the chosen cluster. Both use the PyTorch Geometric framework to run the deep learning process, and both are natively built to make use of GPUs, if the underlying system offers them. Custom loss functions and a Graph U-Net model were customised for this task.
Next versions will improve the prediction capabilities of the trained network. Improvements will include a data augmentation step for the generation of sub-regions to be used to improve the generalisation capabilities of the GNN, as well as a final post-processing step to test the network on a bigger region. Other improvements aim at improving the prediction results. Unfortunately, it’s not possible to use future projection data like CMIP6 due to its resolution not being high enough for the detection of oceanic mesoscale eddies. However, it will be possible to have a tool that can produce segmentation masks directly on unstructured climate grids and do it faster than the classic physics-based algorithms that only work on regular matrix data. A multi-model (both for data and ML models) thematic module could be also explored to support ensemble detection approaches.