Aichinger-Rosenberger & Sjoberg (2026): Retrieval of thermodynamic profiles in the lower atmosphere from GNSS radio occultation using explainable deep learning: Model, code and validation data

This repository contains the assets to the revised GMD manuscript Aichinger-Rosenberger & Sjoberg (2026): Retrieval of thermodynamic profiles in the lower atmosphere from GNSS radio occultation using explainable deep learning. It contains the new AROMA model and related scaling routines (model.zip), the python scripts to run the whole pipeline (code.zip), additional files produced or needed by the specific routines as well as the radiosonde data used for validation (raob.zip). model.zip: contains the trained model (mlp_best.pt) as well as the scaling routines for features and targets (in the folder datasets) code.zip: contains script for the whole AROMA processing pipeline + some additional files read/written by different scripts The exact pipeline looks like this: write_wetPf3.py interpolate_era5_wetPf3.py aroma_wetPf3_zarr.py aroma_preprocess.py aroma_train.py aroma_inference.py Then various types of validation with visualizations: aroma_validation_testdata.py aroma_validation_ERA5.py aroma_validation_raob.py aroma_validation_testdata.py aroma_validation_raob_profiles.py Additionally, it includes helper functions (aroma_functions.py) and the actual model (aroma_models.py) which are needed in the same directory to be able to run the other scripts. raob.zip: Since the radiosonde data does not have a permanent DOI, we provide the data here in this folder. In addition, it contains text files providing information on all collocated profile pairs (RO and radiosondes). All other data (RO and ERA5) is publicly available from the respective providers cited in the paper (UCAR and Copernicus). Furthermore their inclusion here is not possible due to their size (>> 100 GB total).