Foehn: trained model weights (global kilometre-scale forecasts of fine particulate matter)
Authors/Creators
Description
Trained weights of Foehn, the generative model of the manuscript "Global kilometre-scale forecasts of fine particulate matter" (Kheder, Boy and Liu, submitted). The archive foehn_weights.zip contains five checkpoints: foehn_1km_final.ckpt, the model of the paper (14.2 million parameters; all main figures and Extended Data Figs. 5, 6 and 9); foehn_1km_before_crps_training.ckpt, the same model before the CRPS calibration (Extended Data Fig. 6h); foehn_1km_base_no_emissions.ckpt, the base model trained without the six emission channels (Fig. 5j and Extended Data Fig. 9); foehn_1km_no_transport.ckpt, the same network without the computed transport (Fig. 5j); and foehn_cams_aurora_protocol.ckpt, Foehn trained in Aurora's protocol on the CAMS analyses (9.9 million parameters; Extended Data Figs. 7 and 8). The README gives the correspondence between files and figures, the SHA-256 checksums, the exact commit of the code (https://github.com/AmmarKheder/Foehn, commit b523bc3), the software environment and the inference commands for the 2022 evaluation, for a single start day and for each model. Licence: MIT. Access to the files is restricted during peer review and will be opened upon publication.
Files
Additional details
Software
- Repository URL
- https://github.com/AmmarKheder/Foehn