IONIS Sovereign AI Lab¶
"The logs have been speaking for decades, but nobody is listening."
Every day, millions of radio contacts are logged — WSPR beacons, RBN spots, contest QSOs, PSK Reporter decodes. They contain the ground truth of HF propagation: what actually worked, when, and under what conditions. Until now, that data sat in archives, unused for prediction.
What is IONIS?¶
IONIS (Ionospheric Neural Inference System) is a machine learning engine that predicts HF radio propagation using real-world observations instead of theoretical models.
IONIS answers one question: "Can I work this path right now, on my mode?"
One model predicts the signal level. A mode-aware threshold layer converts that into operational verdicts for six mode families — from WSPR at -28 dB to SSB at +5 dB. One forward pass, six answers.
Building on Traditional Prediction¶
Tools like VOACAP represent decades of ionospheric research and remain valuable references. But they have inherent limitations:
- Static models: Based on historical averages, updated infrequently
- Single-mode focus: Designed for SSB voice circuits, with no concept of digital mode thresholds
- Limited validation: No feedback loop — predictions are never checked against observations
- Coarse resolution: Monthly medians miss daily and hourly variations
When FT8 operators use VOACAP and find "closed" paths that are wide open at -20 dB, that's not a VOACAP bug — it was never designed for the digital world.
IONIS extends VOACAP's legacy by adding what it lacks: mode-aware thresholds, continuous learning from real-world observations, and a closed-loop feedback cycle. The two approaches validate each other — physics models provide theoretical grounding, while observational data reveals what actually happens.
The IONIS Approach¶
IONIS is built on four pillars:
- Massive observational data: 13.18B+ radio contacts from four independent networks (WSPR, RBN, contest logs, and PSK Reporter)
- Neural network with physics constraints: The model can't violate known ionospheric physics
- Mode-aware prediction: One SNR prediction yields six operational verdicts (WSPR, FT8, CW, RTTY, SSB)
- Closed-loop validation: Live PSK Reporter data continuously scores the model against observations it has never seen
The model predicts signal-to-noise ratio — a physical quantity that is mode-agnostic. The threshold layer applies mode-specific decode limits to determine operational viability. The physics doesn't change by mode; only the minimum signal required to use it does.
Current Status¶
IONIS V20 Golden Master is the production model. Trained on 20M WSPR + 4.55M DXpedition (50x) + 6.34M Contest signatures (~31M rows). It correctly predicts:
- Higher solar flux (SFI) improves propagation (+0.482σ, ~3.2 dB benefit, monotonic)
- Geomagnetic storms (Kp) degrade propagation (+3.487σ, ~23.4 dB cost, monotonic)
- Path geometry, time of day, and seasonal effects
| Metric | Value |
|---|---|
| Pearson correlation | +0.4879 |
| RMSE | 0.862σ (~5.8 dB) |
| Recall (vs VOACAP) | 96.38% (+20.56 pp vs VOACAP) |
| PSK Reporter Recall | 84.14% (independent live data) |
| Physics Tests | 4/4 PASS |
V20 demonstrates consistent improvement over the ITS/NTIA reference model (VOACAP) on 1M validated contest paths. PSK Reporter live validation confirms the model generalizes to data it has never seen. For digital modes (FT8, FT4, WSPR) and CW, IONIS provides predictions where no comparable reference model exists.
Data Sources¶
| Source | Volume | Mode Coverage | Purpose |
|---|---|---|---|
| WSPR | 10.8B spots | WSPR (-28 dB) | Signal floor, continuous baseline |
| RBN | 2.18B spots | CW, RTTY | Machine-decoded SNR, DXpedition coverage |
| Contest Logs | 195M QSOs | SSB, CW, RTTY, Digi | Ground truth — proof the band was open |
| PSK Reporter | ~26M/day (live) | FT8, FT4, WSPR | Real-time validation and future training |
| Solar Indices | 76K rows | — | SFI, Kp, SSN conditions (2000-2026) |
Infrastructure¶
Self-hosted, no cloud dependencies:
| Host | Role |
|---|---|
| Threadripper 9975WX | Control node — ClickHouse, CUDA, Go pipelines |
| Mac Studio M3 Ultra | Sage node — PyTorch training |
| EPYC 7302P | Forge node — backup/replica |
Documentation¶
- Architecture — IonisGate model design, sidecars, gated physics
- Methodology — Data ingestion and training process
- Validation — 35-test automated verification suite
- Data Privacy — GDPR compliance, data authorization
- Ethos — Core principles and statement of intent
Open Source¶
All code, data pipelines, and documentation are open source. The goal is a reproducible, verifiable propagation prediction system that anyone can run.
Built by KI7MT. No cloud. No vendor lock-in. Just physics and data.