Research
Pre-trained models
for orbital ML
Models for orbit prediction, conjunction analysis, and distributed space computing. Trained on public data, Apache 2.0 licensed.
Our approach
We train models on publicly available data and release them for the research community.
Public training data
Trained on TLEs from Space-Track, space weather from NOAA, and our own simulation experiments. No proprietary data required.
Reproducible
Training code, data processing scripts, and evaluation protocols available. Verify our results and build upon them.
Open license
Apache 2.0 for all models. Use commercially, modify freely, no attribution required.
Orbital Intelligence Models
Models for satellite tracking, conjunction analysis, and space situational awareness. Trained on public Two-Line Element data.
orbml-base
General-purpose orbit prediction model. Given historical TLE observations, predicts future orbital elements with uncertainty estimates. Trained on millions of TLEs from Space-Track.
conjunctionnet
Collision probability estimation model. Takes two objects' orbital states and covariances, outputs probability of collision and time of closest approach. Trained on computed conjunction events from our dataset.
maneuver-detect
Satellite maneuver detection and classification. Analyzes TLE sequences to identify when maneuvers occurred and classify type (station-keeping, orbit raise, plane change, collision avoidance).
satclass
Satellite classification from orbital behavior. Given only orbital elements over time, classifies satellite type (communication, Earth observation, navigation, scientific) and operational status.
reentry-predict
Atmospheric reentry prediction. Combines orbital mechanics with atmospheric density models to predict reentry time windows for decaying objects. Trained on historical reentry events.
Distributed Compute Models
Models for federated learning, model partitioning, and coordination across distributed infrastructure with space-like constraints.
gradient-compress
Learned gradient compression for bandwidth-limited federated learning. Compresses gradients during distributed training with minimal accuracy loss. Applicable to any bandwidth-constrained distributed training scenario.
model-partition
Neural network partitioning optimizer. Determines optimal layer placement across distributed nodes given latency, bandwidth, and compute constraints. Trained on exhaustive partitioning evaluations across model architectures.
sync-scheduler
Data synchronization scheduler for intermittent connectivity. Prioritizes what to sync given limited bandwidth windows, data freshness requirements, and upcoming connectivity. Trained on simulated ground station pass schedules.
bandwidth-predict
Link capacity predictor for ground station passes. Estimates achievable throughput based on orbital geometry, elevation profile, and atmospheric conditions. Trained on link budget models and atmospheric data.
How to use
Via API
Access models through our REST API. No setup required - just send requests and get predictions.
Via SDK
Use our Python, Node.js, or Rust SDKs for native integration. Handles auth, retries, and batching.
Download weights
Download model weights for self-hosted deployment. ONNX and PyTorch formats available.
Model overview
Choose the right model for your use case.
Orbital Intelligence Models
| Model | Task | Training Data | Best for |
|---|---|---|---|
| orbml-base | Orbit prediction | Space-Track TLEs | General orbit propagation |
| conjunctionnet | Collision probability | Conjunction Events Dataset | Conjunction screening |
| maneuver-detect | Maneuver detection | Maneuver Detection Dataset | Behavior analysis |
| satclass | Classification | Satellite Classification Dataset | Catalog enrichment |
| reentry-predict | Reentry prediction | Historical reentries | Debris tracking |
Distributed Compute Models
| Model | Task | Training Data | Best for |
|---|---|---|---|
| gradient-compress | Gradient compression | FL Experiment Logs | Bandwidth-limited FL |
| model-partition | Model splitting | Partitioning Results | Distributed inference |
| sync-scheduler | Data synchronization | GS Visibility Dataset | Intermittent connectivity |
| bandwidth-predict | Link capacity | Link budget models | Transfer planning |
Launching Q2 2026
We're finalizing model training and evaluation. Contact us for early access or research collaboration. Benchmarks and evaluation results will be published alongside model release.
Want early access to models?
Contact us if you're working on space research and want to evaluate our models before public release.