Documentation site for the IONIS (Ionospheric Neural Inference System) project.
Built with MkDocs Material.
IONIS is a machine learning system that predicts HF radio propagation using real-world observations. Instead of relying solely on theoretical ionospheric models, IONIS learns from billions of heterogeneous observations (WSPR, RBN, Contest Logs) harmonized through a Z-normalized signal-to-noise architecture.
Current Model: IONIS V20 — Production
- IonisGate architecture (203,573 parameters)
- Trained on 20M WSPR + 4.55M DXpedition (50x) + 6.34M Contest = ~31M rows
- Pearson +0.4879, RMSE 0.862σ, Kp +3.487σ, SFI +0.482σ
- Step I Recall: 96.38% vs VOACAP 75.82% (+20.56 pp)
- PSK Reporter live validation: 84.14% recall on independent data
- 100 epochs in 4h 16m on Mac Studio M3 Ultra (MPS backend)
git clone https://github.com/IONIS-AI/ionis-docs.git
cd ionis-docs
pip install -r requirements.txt
mkdocs serveThe site will be available at http://127.0.0.1:8000/.
| Repository | Purpose |
|---|---|
| ionis-training | PyTorch model training and validation |
| ionis-apps | Go data ingesters (WSPR, solar, contest, RBN) |
| ionis-core | DDL schemas, SQL scripts, base configuration |
| ionis-cuda | CUDA signature embedding engine |
GPLv3 — See LICENSE for details.
Greg Beam, KI7MT
Built for amateur radio operators who want propagation predictions based on what actually happened, not what theory says should happen.