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 V22-gamma + PhysicsOverrideLayer — Production (Phase 4.0)
- IonisGate architecture (205,621 parameters, 17 features)
- Trained on 38.7M rows: WSPR + DXpedition (50x) + Contest
- Pearson +0.492, RMSE 0.821σ, KI7MT 17/17, TST-900 9/11
- PhysicsOverrideLayer: deterministic high-band night closure clamp
- Checkpoint format: safetensors (805 KB)
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-validate | Model validation suite (PyPI) |
| ionis-apps | Go data ingesters (WSPR, solar, contest, RBN, PSKR) |
| ionis-core | DDL schemas, SQL scripts, base configuration |
| ionis-cuda | CUDA signature embedding engine |
| ionis-hamstats | ham-stats.com publishing |
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.