Intelligence

Constellation is an operational learning system. Every simulation, prediction, decision, and observed outcome feeds the digital twin — continuously improving how future missions are planned and executed, with every forecast bounded by physics and every model promoted with the discipline of flight software.

Philosophy

Physics decides what is possible. ML decides what is likely.

Orbital mechanics, link budgets, and power systems are deterministic — we simulate them rigorously, with ITU-R propagation models and full orbit propagation. What physics can't tell you is how a specific radio degrades in rain, how demand shifts on a Tuesday evening, or which station to pick when a weather cell is moving in.

That's where learned models earn their place: scoped to exactly the gap physics leaves open, trained on fleet telemetry, and never allowed to contradict the link budget — because they never see anything else to learn. A prediction that violates physics isn't filtered out; it can't be produced.

Deterministic physicsOrbit propagationLink budgetsPower & eclipseContact geometryAttitude & constraintsWhat is possibleLearned modelsLink fade forecastingAnomaly detectionTraffic & demandModel calibrationWhat is likelyPhysics-informed MLForecasts with physical bounds — uncertainty you can act on
Physics-informed ML — bounded predictions, not black boxes

Residual learning

The model only learns what physics can't explain

The clear-sky link budget is computed identically in training and serving from ephemeris and RF configuration. Models are trained on the residual — the difference between observed link quality and what physics predicts — and forecasts are reconstructed by adding the physics back at each horizon.

Clear-sky link budgetSNR = EIRP + G/T − FSPL(d, f)+ 228.6 − 10·log₁₀(B)from ephemeris + RF config · deterministicObserved telemetrySNR as measured on the linkResidual r(t)what physics can't explainModelforecasts r̂ ateach horizonforecast(t+h) = clear-sky physics(t+h) + r̂ — a prediction that violates the link budget cannot exist
Physics-informed by construction — the model only learns the residual

The model portfolio

Forecasters for links, demand, and decisions

Link-quality forecasting

Forecasting

Where every link is heading over the next minutes — as a full distribution with explicit fade and usability probabilities, not a point estimate.

Demand & capacity forecasting

Forecasting

Where traffic is going, not where it was — so capacity decisions run ahead of demand instead of behind it.

Station selection

Decision support

Which ground station to use for every pass, scored across the entire federated network against mission intent.

Physics baseline

Deterministic

A pure-physics forecaster is always on duty. Learned models ship only when they demonstrably beat it — and the system degrades toward physics, never toward silence.

Uncertainty

Distributions, not point estimates

Every forecast ships as an envelope — p10 / p50 / p90 with empirically calibrated coverage — plus explicit probabilities against operational thresholds. A scheduler doesn't need the future; it needs an envelope it can trust. Models are calibrated per site and per fleet, and continuously validated against live operations.

Link SNR (dB)t+1 · t+3 · t+5 minNowUsability thresholdp10p50p90P(below threshold) = 0.12Intervals are conformally calibrated against live residuals and gated to 70–90% empirical coverage before promotion.
Calibrated uncertainty — intervals your scheduler can bet on

Learning across fleets

Why the platform gets smarter than any single operator

Every fleet is different. Every radio ages differently, every modem has its own personality in rain, every optical terminal its own weather sensitivity, every constellation its own geometry and scheduling habits. No single fleet observes every operational regime — which means no single operator's data can teach a model everything that matters.

Physics is shared; residuals are not. The deterministic models are universal. What compounds across the platform is the infrastructure of learning: the residual-learning framework, the feature engineering, the simulation regimes, and the evaluation gates — each made richer by every climate, hardware generation, and failure mode the platform encounters. Every new deployment expands the platform's understanding of where reality consistently diverges from theory.

PhysicsSimulationPredictionsOperationsTelemetryDigital twinResidual learningBetter predictionsBetter operationsevery pass improves the next
The operational flywheel — every decision becomes a labeled example

Every operational decision — successful or unsuccessful — becomes another labeled example inside the digital twin.

Fleet A — isolated tenantTelemetryFleet twinFleet-specific modelscalibrated on your telemetryFleet B — isolated tenantTelemetryFleet twinFleet-specific modelscalibrated on your telemetryFleet C — isolated tenantTelemetryFleet twinFleet-specific modelscalibrated on your telemetryShared learning infrastructurephysics engine · simulation regimes · residual-learning framework · feature engineering · evaluation gatesWhat flows down is capability, not data. Models improve per fleet; the framework improves for everyone.Physics is shared. Residuals are not.
Shared infrastructure, never shared telemetry

The isolation guarantee

Fleet-specific models remain calibrated on each customer's telemetry, inside that customer's enclave. Statistical improvements to the residual-learning framework, feature engineering, and simulation regimes benefit the platform as a whole — while preserving customer isolation. Shared infrastructure, never shared telemetry.

Where rare events come from

Deep rain fades, gateway outages, solar events, mass conjunctions, hardware failures — the events operators care about most are the ones fleets almost never experience. The simulation engine generates these regimes at ITU-R fidelity, so models learn from thousands of physically realistic worst cases before meeting a real one.

Constellation doesn't become more valuable because it stores more telemetry. It becomes more valuable because every mission expands the operational knowledge encoded in its physics, simulation, and decision engines — the learning layer for space operations.

Why both

Deterministic simulation and learned models are stronger together

Physics alone

Exact about geometry, power, and links

Blind to weather dynamics, hardware aging, and real traffic

ML alone

Adapts to messy, real-world data

Unbounded — free to predict what physics forbids

Together

Learned models constrained to the residual physics leaves open

Predictions stay inside physical reality — and improve every pass

From prediction to action

A forecast that doesn't change the schedule is just trivia

Predictions flow into the same control plane that schedules contacts, selects ground stations, and routes traffic. Operator policy defines the envelope; inside it, actions execute automatically and every outcome feeds back into the models. This is the loop behind every product in the suite.

Telemetrylive fleet stateForecastphysics + MLPolicy checkoperator envelopeActionschedule · routeFleetverified by telemetryOutcomes retrain the models — every pass improves the next
From prediction to operational action — a closed loop

The full architecture — training pipelines, serving topology, and the security model around them — is documented in the technical whitepaper.

Starta30-dayshadowpilot

Engage

For mission operations teams

Review your ground data architecture on a discovery call.

Book a discovery call

Step 1 of 4

Questions? contact@constellation.space

Technical whitepaper

Enter your email to access the ConstellationOS technical whitepaper.