OrbitGuard AI

The SGP4 propagator converts TLE data into position and velocity vectors at any given epoch, accounting for perturbations from Earth's oblateness (J2), atmospheric drag, and lunar/solar gravitational effects.

Using Runge-Kutta 4th order numerical integration with adaptive step-size control, we propagate state vectors forward. The propagator accounts for J2-J6 zonal harmonics, solar radiation pressure, and third-body perturbations.

A spatial filter first eliminates objects with non-overlapping orbital shells. Then a temporal filter checks for co-location within the screening window. Smart box filtering reduces O(n^2) to O(n log n) complexity.

Transform relative position vector from ECI to RTN frame centered on primary object. Apply iterative refinement using Newton-Raphson method to find exact TCA to microsecond precision.

State transition matrices propagate 6x6 covariance matrices from epoch to TCA. The combined covariance is projected onto the encounter plane perpendicular to relative velocity vector.

The 2D Pc integral assumes linear relative motion and Gaussian uncertainty. For high-eccentricity or long-duration encounters, we switch to Chan's 3D method or Monte Carlo sampling.

A trained deep ensemble model ingests 47 conjunction features (miss distance, Pc, covariance realism metrics, data quality flags) and outputs risk score with epistemic uncertainty via MC Dropout.

Each sample perturbs initial states according to covariance, propagates independently, and checks for collision. Convergence is monitored via running mean and variance to ensure statistical reliability.

Sequential Quadratic Programming (SQP) minimizes delta-V subject to constraints: minimum post-maneuver miss distance, fuel budget, operational windows, and secondary conjunction screening.

Outputs include: burn start time (UTC), delta-V vector in RTN, predicted fuel consumption, post-maneuver orbital elements, and 7-day secondary screening results.