Appearance
New machine learning models for THEMIS ASI and TREx RGB
September 21, 2026
We are pleased to announce a new set of machine learning classification models for our all-sky imager data. This release includes an updated second version of our THEMIS ASI models, and the first release of models for the TREx RGB network.
Each model classifies individual all-sky images in one of two ways:
- Cloud classification - Is cloud visible in the images?
- APA classification - Is amorphous pulsating aurora (APA) visible in the images?
Every classification is based on a 10-minute window, and a confidence value can be used when filtering.
What is new
TREx RGB v1 models - Cloud and APA classification models are now available for the TREx RGB network for the first time. In our internal evaluations, the TREx RGB models perform on par with the THEMIS ASI v2 models.
THEMIS ASI v2 models - Both the cloud and APA models have been retrained and re-evaluated. The v1 models were already performing well, so the improvements are modest for cloud classification and more noticeable for APA classification. The v1 models remain available for anyone who has built analyses around them.
Using the models in AuroraX
The classifications are attached to the ephemeris records in AuroraX as metadata, which means they can be used as filters when searching. In both Conjunction Search and ephemeris search, add a metadata filter on the appropriate field to restrict results to, for example, clear skies with aurora present. The following metadata fields are available:
| Network | Model | Metadata field | Confidence field |
|---|---|---|---|
| THEMIS ASI | Cloud v1 | ucalgary_themis_cloud_ml_v1 | ucalgary_themis_cloud_ml_v1_confidence |
| THEMIS ASI | APA v1 | ucalgary_themis_apa_ml_v1 | ucalgary_themis_apa_ml_v1_confidence |
| THEMIS ASI | Cloud v2 | ucalgary_themis_cloud_ml_v2 | ucalgary_themis_cloud_ml_v2_confidence |
| THEMIS ASI | APA v2 | ucalgary_themis_apa_ml_v2 | ucalgary_themis_apa_ml_v2_confidence |
| TREx RGB | Cloud v1 | ucalgary_trexrgb_cloud_ml_v1 | ucalgary_trexrgb_cloud_ml_v1_confidence |
| TREx RGB | APA v1 | ucalgary_trexrgb_apa_ml_v1 | ucalgary_trexrgb_apa_ml_v1_confidence |
The same filters work in PyAuroraX and IDL-AuroraX. The AuroraX documentation has a dedicated machine learning section describing each model and how to use them in searches. For worked code examples, see the machine learning enhanced searching crib sheets for Python and IDL.
Classifying new data
Classifications for newly acquired data will be added to AuroraX periodically throughout the year, and we will post a bulletin each time an update occurs. The models run on the raw imaging data, not on the real-time products, so there is a natural lag between an observation and its classification that depends on when the raw data from each site is ingested into our archive. The new raw_data_exists metadata field described in our automated data loading news post is a good indicator: records without raw data cannot have classifications yet.
Downloadable ASCII files
New for this release, the full classification output from every model is also available as ASCII files. If you like good 'ol fashioned text files, they're available for you to utilize as you need.
- https://data.aurorax.space/data/ml_models/ - all models in one place
- THEMIS ASI machine learning - on the main data archive
- TREx RGB machine learning - on the main data archive
The files are free to use. If you use them in a publication, we ask that you cite AuroraX as described in the README in each model directory, and see our how to cite page for the relevant instrument acknowledgements.