Deep Learning for Space Object Density Distribution Prediction

Abstract

The rapid growth of artificial objects in Low Earth Orbit (LEO) has heightened concerns over orbital congestion and collision cascades, known as Kessler Syndrome. Traditional high-fidelity models, while accurate, are computationally intensive and poorly scalable. This thesis introduces a machine learning–based framework for forecasting the long-term evolution of space object density. A large dataset is generated, using the MIT Orbital Capacity Assessment Tool – Monte Carlo (MOCAT-MC), simulating thousands of scenarios across varying launch, disposal, and maneuver parameters. A Convolutional Gated Recurrent Unit (ConvGRU) is trained to predict density distributions over a 100-year horizon, achieving accurate forecasts with significantly reduced runtime. With a simple guidance mechanism, the generalization capability of the model across diverse scenarios is greatly improved. This approach offers a scalable and efficient tool for supporting future space traffic management and sustainability efforts.