Virtual EVE predicts solar extreme ultraviolet (EUV) irradiance using deep learning applied to imagery from NASA's Solar Dynamics Observatory (SDO). Developed during FDL-X 2023, the model reconstructs measurements originally provided by the SDO/EVE MEGS-A instrument (which failed in 2014) by predicting all 38 EVE ion spectral lines from 9-channel AIA solar images.
The model uses a hybrid approach that blends:
- A linear model for statistical baseline predictions
- A CNN (EfficientNet-B5 backbone) for learning spatial features from 512x512 px AIA images
- A learnable blending parameter that combines both outputs
Training and experiment tracking use PyTorch Lightning and Weights & Biases.
| Directory | Description |
|---|---|
src/irradiance/ |
Core library — data loading, model definitions, training |
web_app/ |
Streamlit application for interactive irradiance visualization |
inference-cloud-function/ |
Google Cloud Function for serverless inference |
notebooks/ |
Analysis and experimentation notebooks |
The model is trained on the SDOML v2 dataset (Zarr format), with support for both local filesystem and AWS S3 backends. HMI magnetogram data is also integrated into the pipeline.
pip install -e .To run the web app:
cd web_app
streamlit run main.pyFDL-X 2023 ARD EUV Team