deinal/spacecast-models

Models Pretrained on the Vlasiator Dataset for Ion-kinetic Plasma Emulation

The models have been produced using spacecast. The repository contains the following:

model_weights/
├── forecasts/                    - Directory containing example forecasts for each run
├── metrics/                      - Evaluation metrics for each model variation in csv format
├── plots/                        - Evaluation figures showcasing RMSE, CRPS, SSR scores and qualitative results
├── graph_efm_simple.ckpt         - Pretrained Graph-EFM model with simple graph, 12 proc layers and 256 latent dim
├── graph_efm_multiscale.ckpt     - Pretrained Graph-EFM model with multiscale graph, 12 proc layers and 256 latent dim
├── graph_efm_hierarchical.ckpt   - Pretrained Graph-EFM model with hierarchical graph, 6 proc layers and 256 latent dim
├── graph_fm_simple.ckpt          - Pretrained Graph-FM model with simple graph, 12 proc layers and 256 latent dim
├── graph_fm_multiscale.ckpt      - Pretrained Graph-FM model with multiscale graph, 12 proc layers and 256 latent dim
└── graph_fm_hierarchical.ckpt    - Pretrained Graph-FM model with hierarchical graph, 6 proc layers and 256 latent dim

In the forecast directory you will find 50 second forecasts for each of the for runs in spacecast-data. The models used to produce these are graph_efm_simple.ckpt and graph_fm_simple.ckpt. Test set metrics are available in metrics. For a full evaluation including visualization of the example forecasts see the plots directory.