Autonomous orbital optimization
Hours-long orbital simulations are replaced with a learned model, which makes searching satellite layouts practical. The best layouts cut the mean wait between passes over eight disaster regions by 43.9% against the standard 48-hour single-satellite benchmark.
shorter revisit gap than the standard 48-hour single-satellite benchmark
on held-out simulations
of predictions pass the physics check
One high-fidelity simulation takes hours, far too slow to search with. A stacked model trained on 7,000+ of them stands in for the simulator, and scores almost the same on unseen runs (0.892 R²) as in cross-validation (0.897).
A prediction that breaks Kepler's third law or falls outside bounds measured from the simulations is thrown out, so the search cannot exploit a model mistake.
A Gaussian process on the model's own past errors gives each prediction a 95% interval, and the least certain orbits are suggested as the next ones to simulate.
Pick a region
Each dot is one non-dominated orbit design from my NSGA-II search: 310 across 5 regions. Gap is the surrogate model's cloud-adjusted revisit gap in hours. Signal loss is (mean altitude / 400 km)². Model results, not flown orbits.
| Region | Designs | Shortest gap (h) | Lowest loss |
|---|---|---|---|
| Amazon | 70 | 17 | 0.25 |
| Caribbean | 69 | 19.2 | 0.25 |
| Australia | 66 | 37.3 | 0.25 |
| California | 50 | 21.7 | 0.25 |
| United Kingdom | 55 | 18.6 | 0.25 |
Limits. No orbit from this system has been flown. The model is checked against simulation output, not measurements from orbit.
