Code for training the self supervised deep learning model, its final checkpoint, and ICON-EU forward simulated BT forecast data used in the study "Satellite-trained latent spaces for evaluating km-scale atmospheric simulations"

This archive accompanies the manuscript “Satellite-trained latent spaces for evaluating km-scale atmospheric simulations.” It contains the model, code, observationally derived reference information, and ICON-EU data used in the study.

The archive includes:





the code used to train the self-supervised vision-transformer encoder on Meteosat Second Generation (MSG) SEVIRI 10.8-µm brightness-temperature imagery;




The final trained model checkpoint used to extract the 384-dimensional latent representations analyzed in the study;




The ten initial observation-derived cluster centroids obtained before the merging procedure that produced the final eight thermal-scene regimes;




The t-SNE code used to fit the two-dimensional visualization to the observational embeddings as the reference and subsequently transform the IFS experiment i4ql and ICON-EU embeddings into the same fitted map; and




The MSG-equivalent, forward-simulated 10.8-µm brightness-temperature forecast fields from ICON-EU were used in the model evaluation.



In the accompanying study, the trained encoder defines a fixed observational latent space in which satellite observations and simulated brightness-temperature scenes are compared. ... mehrThe initial observational partition contains ten clusters, three of which are subsequently merged into a single deeper-cloud regime, yielding the eight regimes used in the final analysis. These regimes describe differences in cloud organization, cloud-top temperature, clear-sky fraction, and land-surface heating.

The archived t-SNE implementation provides a common two-dimensional visualization by fitting the mapping to the observational latent representations and projecting the IFS and ICON-EU representations into that observationally defined map. The t-SNE representation is used for visualization only; the clustering, regime assignment, and quantitative analyses reported in the study are conducted in the original 384-dimensional latent space.

The original MSG observations are not redistributed in this archive and must be obtained separately from the EUMETSAT Data Store. The IFS data analyzed in the study are also not included because they are subject to ECMWF access and redistribution restrictions. The ground-based cloud-radar, ERA5, and CLAAS-3 data used for supplementary analyses remain available from their respective providers. Access information and formal citations for all external datasets are provided in the Open Research section of the accompanying manuscript.