Physics-informed machine learning for satellite operations

The failure mode we designed against

An end-to-end learned forecaster trained on link telemetry will happily learn the geometry of your constellation — badly. It will memorize that links are usually good at high elevation, encode it in millions of parameters, and then extrapolate confidently into configurations it never saw: a new ground site, a different antenna, an orbit raise. The failure isn't noise; it's a model spending its capacity re-deriving Newtonian mechanics from samples, with none of the guarantees.

But orbital geometry and RF propagation are not statistical questions. Elevation, slant range, free-space path loss — these are computable to high precision from ephemeris and hardware configuration. Learning them from data is strictly worse than computing them.

Residual learning: subtract physics first

Our link-quality models never see raw SNR as a target. The clear-sky link budget is computed deterministically — identically in training and serving — from ephemeris and RF configuration:

SNR_cs = EIRP + G/T − FSPL(d, f) + 228.6 − 10·log10(B)

target:   r(t)   = SNR_observed(t) − SNR_cs(t)
serving:  ŷ(t+h) = SNR_cs(t+h) + r̂(t+h)

The model is trained on the residual r(t) — the part of link behavior physics cannot explain: rain cells, scintillation, hardware aging, interference. At serving time the deterministic term is added back at each horizon. Two properties fall out structurally:

• A forecast that violates the link budget cannot be produced. The physics term is not a filter applied after the model; it is the coordinate system the model lives in.
• The learned component stays small and inspectable. When the model and reality diverge, the residual shows it immediately — and that divergence is itself the anomaly signal and the next round of training data.

Features are physics too

The feature contract — 29 signals per link at 15-second cadence — is dominated by physical quantities: orbital geometry (elevation, azimuth, slant range, elevation rate, time to horizon), atmosphere (rain rate and trend, rain and tropospheric attenuation, humidity, pressure), space weather (Kp, F10.7 solar flux, sun angle, sun noise), and RF configuration (frequency, bandwidth, G/T, EIRP, antenna geometry). One deliberate trick: the model also receives known-future covariates — ephemeris geometry at each forecast horizon. Where the satellite will be is the one thing about the future you are allowed to know, so the model conditions on it explicitly instead of inferring it.

The contract is versioned and validated: feature names, ordering, cadence, and fill policies are asserted identical between the training artifact and the serving path. Drift fails the build, not the forecast.

Simulation closes the data gap

The events that dominate operational risk — deep fades, storm passages, solar activity — are rare in any single fleet's telemetry. Our physics engine synthesizes them at ITU-R fidelity (P.838/P.839 rain attenuation, P.1853 time-series dynamics, full link budgets over SGP4-propagated geometry), giving models training exposure to regimes real fleets encounter rarely but catastrophically. Real telemetry then calibrates what simulation cannot know: the specific behavior of your hardware.

The physics baseline is the bar

This is the discipline that keeps “ML-native” from becoming “ML-shaped marketing”: sealed time-ordered splits, regime-balanced training, evaluation slices declared before the experiment, promotion gates on accuracy, coverage, and latency, canary rollout with bake time, and live scoring that seeds the next challenger. The champion stays until beaten.

Physics decides what is possible. ML decides what is likely. Keeping those roles separate — structurally, not procedurally — is what makes the intelligence trustworthy enough to act on.