IONIS Model¶

IONIS (Ionospheric Neural Inference System) predicts HF radio signal-to-noise ratio (SNR) for any transmitter-receiver path. One model, one forward pass, six mode-aware operational verdicts — from WSPR at -28 dB to SSB at +5 dB.

Current Production: IONIS V20¶

Metric	Value
Architecture	IonisGate (203,573 parameters)
Pearson correlation	+0.4879
RMSE	0.862 sigma (~5.8 dB)
Recall vs VOACAP	96.38% (+20.56 pp)
PSK Reporter recall	84.14% (independent live data)
Physics tests	4/4 PASS

Trained on 20M WSPR + 4.55M DXpedition (50x) + 6.34M Contest signatures (~31M rows) on Mac Studio M3 Ultra. 100 epochs in 4h 16m.

Architecture¶

IonisGate separates geography from physics by design:

Trunk DNN: 11 geography/time features through 512 - 256 - 128 - 1
Sun Sidecar: SFI (solar flux) through a monotonic MLP — higher SFI always helps
Storm Sidecar: Kp (geomagnetic) through a monotonic MLP — storms always hurt
Gated mixing: Trunk-derived gates scale sidecar contributions by geography

The "Nuclear Option" — the trunk receives zero direct solar information. All physics flows through constrained sidecars that cannot violate ionospheric law.

Methodology¶

The training pipeline transforms 13B+ raw radio observations into model-ready signatures through a medallion architecture:

Bronze: Raw ingest from WSPR, RBN, contest logs, PSK Reporter
Silver: CUDA-accelerated embeddings with solar enrichment
Gold: Aggregated signatures — grid-pair, band, time, solar, SNR

Validation¶

V20 is validated against three independent benchmarks:

62-test automated battery — physics, canonical paths, adversarial inputs, regression
1M contest path comparison — IONIS 96.38% vs VOACAP 75.82%
PSK Reporter acid test — 84.14% recall on data the model has never seen