Skip to content

From Empirical Physics to Neural Prediction

While legacy tools like VOACAP rely on empirical formulas derived from 1960s ionospheric measurement campaigns, IONIS uses physics-constrained PyTorch neural networks to predict HF path viability from first principles of observed behavior. Trained on over 14 billion real amateur radio observations from WSPR, the Reverse Beacon Network, and contest logs, IONIS captures nonlinear ionospheric behaviors — sporadic-E openings, real-time solar variability, grey-line enhancement — that traditional mathematical models miss. On 1 million real contest paths, IONIS achieves 96.38% recall versus VOACAP's 75.82%.

VOACAP (Voice of America Coverage Analysis Program) has been the standard HF propagation prediction tool since the 1980s. Built on decades of ionospheric research at NTIA/ITS, it uses empirical ionospheric coefficients and ray-tracing geometry to estimate circuit reliability for voice communication.

IONIS takes a fundamentally different approach: learn propagation behavior directly from 14 billion real radio observations, constrained by known physics.

This page documents where we are, what we've proven, and what we're still working on.

Validating Neural Networks vs. VOACAP (V20 Baseline)

IONIS V20 was the first production model validated against both historical contest data and live PSK Reporter observations. The results established that a physics-constrained neural network can outperform empirical models on real amateur radio paths.

Head-to-Head: 1 Million Contest Paths

Both models were given 1M real contest QSOs (CQ WW, CQ WPX, ARRL DX — 2005 to 2025) and asked: "was this band open?" Every QSO actually happened, so the ground truth is always YES.

Model Recall Delta
IONIS V20 96.38% —
VOACAP 0.7.5 75.82% -20.56 pp

The gap was largest on 10m (+38.1 pp) where VOACAP misses sporadic-E and day-to-day solar variability, and on SSB (+22.6 pp) — the mode VOACAP was specifically designed for.

See IONIS vs VOACAP Comparison for full results by mode and band.

Independent Live Validation

V20 achieved 84.14% recall on 100K live PSK Reporter spots the model had never seen during training. This confirmed that IONIS generalizes to independent real-time data, not just historical patterns.

Physics Constraints Hold

The architecture enforces two non-negotiable ionospheric laws through monotonic neural network sidecars:

  • Sun sidecar: Higher solar flux (SFI) always improves propagation (+0.482σ, ~3.2 dB)
  • Storm sidecar: Higher geomagnetic activity (Kp) always degrades propagation (+3.487σ, ~23.4 dB)

These constraints cannot be overridden by the DNN trunk. See Monotonic Sidecars for the full design.

Refining the Deep Learning Architecture: Auroral Zones and Day/Night Physics

V20 proved the architecture works. The ongoing refinement series focuses on giving the model better tools to distinguish when and where propagation behaves differently — auroral zone vulnerability, band-specific day/night behavior, and the separation of storm physics from temporal artifacts.

Modeling Auroral Zone Vulnerability for Polar HF Paths

Problem: V20 treated all paths equally during geomagnetic storms. In reality, a path crossing the auroral zone (high latitude) is far more vulnerable to storm absorption than an equatorial path at the same Kp.

Solution: vertex_lat — the latitude of the highest point on the great circle path. Computed from TX/RX grids via spherical trigonometry:

vertex_lat = arccos(|sin(bearing) * cos(tx_lat)|)

High vertex latitude means the path crosses the polar region where storm particles precipitate. Low vertex latitude means the path stays in the mid-latitudes where storms have minimal effect. This gives the Kp sidecar path-specific context instead of treating storms as a global scalar.

The 10 MHz Pivot: Mathematically Modeling Day and Night HF Bands

Problem: Darkness helps low-band propagation (160m, 80m, 40m) but kills high-band propagation (10m, 15m, 17m, 20m). The crossover is around 10 MHz (30m band). V20 had no way to express this — the model could learn that darkness matters, but not that it matters in opposite directions depending on frequency.

Solution: Solar depression angles at both TX and RX locations, combined with frequency-centered cross-products:

  • tx_solar_dep, rx_solar_dep — continuous day/night indicators (positive = daylight, negative = darkness depth)
  • freq_centered = (freq_mhz - 10.0) / scale — centers frequency around the 10 MHz ionospheric crossover
  • freq_x_tx_dark, freq_x_rx_dark — cross-products that flip sign at the pivot

When the cross-product is positive (high band + darkness, or low band + daylight), propagation is degraded. When negative (high band + daylight, or low band + darkness), propagation is enhanced. The model learns the magnitude; the math enforces the direction.

Isolating Geomagnetic Storm Physics from Temporal Artifacts

An unexpected discovery during V21 training: the V20 storm sidecar's +3.49σ cost was approximately 60% temporal contamination. The sidecar had been compensating for missing time-of-day features — storms correlate with certain hours, and without explicit day/night tools, the Kp sidecar absorbed that temporal signal.

Adding solar depression angles gave the DNN trunk its own time-of-day tools. The Kp sidecar shed its temporal weight and distilled down to pure geomagnetic storm physics. This is the correct behavior: the sidecar should model absorption and fading, not sunrise timing.

The SFI/Kp asymmetry (~3:1 ratio in sigma) is itself physically meaningful. Solar flux acts like a power plant — it lifts the MUF with diminishing returns as ionization saturates. Geomagnetic storms act like a governor — D-layer absorption is violent and nonlinear with no ceiling. Building is bounded; destroying is not.

Scaling Training Data: Integrating 2 Billion CW/RTTY Observations

V17–V19 failed when RBN data (2.18B CW/RTTY spots) was added to training. The post-mortem revealed this wasn't because RBN data was poison — the architectural constraints were incomplete. With the V16 Physics Laws intact plus the frequency pivot, V22 tests whether the "container" is now strong enough to hold the full dataset.

RBN provides critical low-band coverage: 3.18M 160m signatures (nearly 2x WSPR's 1.69M on that band) and 16.3% low-band concentration versus WSPR's 8.1%. If the model needs to learn 160m behavior, it needs RBN data to learn from.

Where VOACAP Still Has Physics We Don't

VOACAP's empirical coefficients encode 60 years of ionospheric measurement campaigns. Some of that physics is not yet available to IONIS:

Path-specific ionospheric state: VOACAP uses ionospheric models to compute the critical frequency (foF2) and peak height (hmF2) at each point along the path. IONIS receives only a global solar flux index — two paths at the same SFI can have wildly different ionospheric conditions depending on latitude, local time, season, and magnetic geometry.

MUF estimation: VOACAP computes the Maximum Usable Frequency for each circuit. IONIS infers this indirectly from frequency, distance, and solar features. A direct MUF signal would tell the model immediately whether the operating frequency can propagate.

D-layer absorption: VOACAP models absorption through the D and E layers explicitly. IONIS handles this through the Kp sidecar globally, without path-specific absorption context.

These gaps point directly to the next phase of model development.

Next: Physics Model Integration

Two open-source tools fill the gaps between what IONIS has learned and what ionospheric physics can provide:

dvoacap-python (pure Python VOACAP port) enables systematic comparison: generate VOACAP predictions for every path in the validation set, then build a diagnostic matrix showing where IONIS succeeds and VOACAP fails, where VOACAP succeeds and IONIS fails, and where both struggle. The paths where VOACAP is right and IONIS is wrong point to known physics we haven't captured yet.

PyIRI (pure Python IRI-2020) provides the path-specific ionospheric features IONIS currently lacks. IRI computes foF2, hmF2, and foE for any location, date, and solar activity level — covering the full training dataset from 1958 to present with no coverage gaps. The foF2/frequency ratio is particularly valuable: it tells the model directly whether the operating frequency is above or below the MUF at each point along the path.

These features would replace the global SFI scalar with a path-specific ionospheric scalpel — giving the model the same physics VOACAP has, plus everything it learns from 14 billion observations that VOACAP has never seen.

The Larger Picture

Empirical models like VOACAP represent the ceiling of what physics alone can predict from averaged ionospheric measurements. Neural networks like IONIS represent what machine learning can extract from billions of individual observations. Neither approach is complete on its own.

The path forward combines both: use physics models to diagnose where the neural network is weak, use ionospheric models to provide features the neural network can't derive from radio data alone, and use the neural network to capture patterns that no empirical model can express — sporadic-E openings, trans-equatorial propagation, grey-line enhancement, and the thousands of subtle effects buried in 14 billion data points.

The logs were speaking for decades. Now we're listening — and teaching a model to understand what they're saying.

Project Resources