Research
Satellite Collision Modeling
Monte Carlo simulation of conjunction risk for 40,000+ tracked objects in low Earth orbit.
Tracked objects
40,000+
Domain
Low Earth orbit
Lab
Duke Cosmology (Scolnic)
01
Overview
Low Earth orbit is increasingly crowded: tens of thousands of satellites and debris fragments share a shell of space where a single collision can cascade into thousands of new hazards. In Professor Daniel Scolnic’s lab at the Duke Cosmology and Astrophysics Research Center, I build tools to quantify that risk.
The core of the project is a Monte Carlo pipeline that ingests public tracking data for roughly 40,000 objects, propagates their orbits forward in time, and estimates collision probabilities between conjunction pairs by sampling from positional uncertainty distributions.
02
Technical approach
- 01
Data ingestion from TLE catalogs
Two-line element sets for the full public catalog are parsed and validated, giving orbital state for every tracked satellite and debris object.
- 02
Orbit propagation with SGP4
Each object’s trajectory is propagated with the SGP4 analytic model, screening for close approaches between candidate pairs over the prediction window.
- 03
Covariance sampling
For each conjunction, positional uncertainty is modeled as a multivariate Gaussian. Thousands of sampled state vectors per object turn a deterministic miss distance into a probability of collision.
- 04
Visualization tools
Python visualization utilities render the full catalog and highlight high-risk conjunction geometry, making results legible to researchers who are not simulation specialists.
03
Architecture
01
TLE Catalog
~40,000 tracked objects
02
SGP4 Propagation
Trajectory prediction
03
Conjunction Screening
Close-approach pairs
04
Monte Carlo Sampling
Covariance-based draws
05
Collision Probability
Risk estimates + visualization
04
Challenges
Scale
Pairwise screening across 40,000 objects is quadratic in principle. Spatial filtering and coarse pre-screens were needed to keep the pipeline tractable before any expensive sampling runs.
Uncertainty modeling
TLE data carries significant positional error that grows with propagation time. Choosing realistic covariance assumptions was as important as the sampling machinery itself.
Rare-event estimation
Collisions are extremely low-probability events, so naive sampling wastes most of its draws. Sample sizes and variance had to be balanced carefully to produce stable estimates.
05
Results
A working end-to-end pipeline from raw TLE data to per-conjunction collision probability estimates.
Visualization tools used within the lab to inspect the full LEO catalog and individual high-risk encounters.
A foundation for the follow-on quantum satellite routing work, which reuses the same TLE infrastructure.
06
Media & links
More projects