← all projects

A Comprehensive Machine Learning Paradigm for Space Debris Surveillance

three machine learning models that find debris, predict its orbit and assess collision risk, and a space based laser design for debris removal

project: 01 of 05
event: ISEF, 4th in Physics & Astronomy
credit: Research project with one partner, Nived Nandakumar. published on SSRN.
links: Read the paper →

01the problem

millions of fragments from old rockets and dead satellites now orbit Earth
Kessler syndrome is the runaway case, where each collision creates debris that causes the next
ground telescopes only track the larger pieces and the atmosphere distorts what they see
no system has actually removed debris, only proposals and planned missions

fragments sharing low Earth orbit with working satellites

Tracked objects in orbit over a 64 year period. Source: European Space Agency, 2024. — Fig. 1. Tracked objects in orbit over a 64 year period. Source: European Space Agency, 2024.

02model one, detection

convolutional neural network trained on over 5,000 labeled orbital images
multi task loss combining classification, localization and confidence
anchor boxes refined against ground truth by minimizing intersection over union
random scaling, rotation and flipping to make it generalize

the network boxes and classifies each object it finds

Detection output on a held out test image. — Fig. 2. Detection output on a held out test image.

Localization loss, penalizing a box in the wrong place or of the wrong size. — Eq. 1. Localization loss, penalizing a box in the wrong place or of the wrong size.

03model two, trajectory

physics informed neural network with an LSTM layer reading trajectory history
Kepler's third law, eccentricity limits, drag and gravity all enforced as loss terms
a plain network drifts into orbits that cannot physically exist, the constraints stop that
time derivatives modeled with gradient tape so predictions stay smooth

faster at perigee, slower at apogee, the way a real orbit moves

Total loss, combining prediction error with four weighted physics penalties. — Eq. 2. Total loss, combining prediction error with four weighted physics penalties.

Kepler's third law written as a residual. Zero when period and semi-major axis agree. — Eq. 3. Kepler's third law written as a residual. Zero when period and semi-major axis agree.

A predicted orbit for a piece of Ariane 42P rocket debris. — Fig. 3. A predicted orbit for a piece of Ariane 42P rocket debris.

04model three, collision risk

random forest regressor, 100 decision trees on bootstrapped samples
works from relative speed, distance and approach angle
risk predicted on a log base 2 scale for resolution at low probabilities
averaging decorrelated trees cuts variance, which the formula makes explicit

inputs fan out to 100 trees and the answers are averaged

Why averaging works, and why decorrelating the trees matters. — Eq. 4. Why averaging works, and why decorrelating the trees matters.

05the laser application

the three models feed a proposed satellite carrying a pulsed laser
focused heat vaporizes a thin surface layer, which is ablation
the plasma plume pushes the fragment onto a lower orbit until it burns up
more precise than any ground based attempt because the platform is already alongside

the ablation plume nudges the fragment onto a lower orbit

06results

detection reached 95.4 percent mean average precision, 98.0 percent on recall confidence
trajectory model reached R squared 0.979 with mean squared error 0.015
collision risk model reached R squared 0.981 with mean squared error 0.014
all three run behind one interface, upload an image and get object, orbit and risk together

Precision against recall for every class. — Fig. 4. Precision against recall for every class.

The interface running detection, orbit prediction and collision scoring on one uploaded image. — Fig. 5. The interface running detection, orbit prediction and collision scoring on one uploaded image.