Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards a Machine Learning-Based Approach to Predict Space Object Density Distributions

Victor Rodriguez-Fernandez, Sumiyajav Sarangerel, Peng Mun Siew, Pablo Machuca, Daniel Jang, Richard Linares

TL;DR

This paper tackles the problem of predicting space object density distributions in LEO as ASO populations grow. It proposes MOCAT-ML, a machine learning extension of MOCAT, to accelerate propagation of phase-space density distributions using three architectures including a 2D autoencoder with ConvGRU forecaster. Results show the end-to-end 2D ConvGRU best for short-term forecasts, achieving rapid inference, but long-term predictions degrade quickly under iterative forecasting. The work highlights hybrid MC-ML strategies and graph-based representations as promising directions to balance accuracy, uncertainty, and computational efficiency for space traffic management.

Abstract

With the rapid increase in the number of Anthropogenic Space Objects (ASOs), Low Earth Orbit (LEO) is facing significant congestion, thereby posing challenges to space operators and risking the viability of the space environment for varied uses. Current models for examining this evolution, while detailed, are computationally demanding. To address these issues, we propose a novel machine learning-based model, as an extension of the MIT Orbital Capacity Tool (MOCAT). This advanced model is designed to accelerate the propagation of ASO density distributions, and it is trained on hundreds of simulations generated by an established and accurate model of the space environment evolution. We study how different deep learning-based solutions can potentially be good candidates for ASO propagation and manage the high-dimensionality of the data. To assess the model's capabilities, we conduct experiments in long term forecasting scenarios (around 100 years), analyze how and why the performance degrades over time, and discuss potential solutions to make this solution better.

Towards a Machine Learning-Based Approach to Predict Space Object Density Distributions

TL;DR

This paper tackles the problem of predicting space object density distributions in LEO as ASO populations grow. It proposes MOCAT-ML, a machine learning extension of MOCAT, to accelerate propagation of phase-space density distributions using three architectures including a 2D autoencoder with ConvGRU forecaster. Results show the end-to-end 2D ConvGRU best for short-term forecasts, achieving rapid inference, but long-term predictions degrade quickly under iterative forecasting. The work highlights hybrid MC-ML strategies and graph-based representations as promising directions to balance accuracy, uncertainty, and computational efficiency for space traffic management.

Abstract

With the rapid increase in the number of Anthropogenic Space Objects (ASOs), Low Earth Orbit (LEO) is facing significant congestion, thereby posing challenges to space operators and risking the viability of the space environment for varied uses. Current models for examining this evolution, while detailed, are computationally demanding. To address these issues, we propose a novel machine learning-based model, as an extension of the MIT Orbital Capacity Tool (MOCAT). This advanced model is designed to accelerate the propagation of ASO density distributions, and it is trained on hundreds of simulations generated by an established and accurate model of the space environment evolution. We study how different deep learning-based solutions can potentially be good candidates for ASO propagation and manage the high-dimensionality of the data. To assess the model's capabilities, we conduct experiments in long term forecasting scenarios (around 100 years), analyze how and why the performance degrades over time, and discuss potential solutions to make this solution better.
Paper Structure (14 sections, 7 figures, 2 tables)

This paper contains 14 sections, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Backgrounds
  3. MIT Orbital Capacity Tool
  4. Deep learning concepts relevant to this work
  5. Dataset
  6. Proposed approaches
  7. Decoupled one-dimensional autoencoder and TST forecaster (Decoupled 1D AE + TST)
  8. End to end one-dimensional autoencoder and LSTM forecaster (e2e 1D AE + LSTM)
  9. End to end bi-dimensional autoencoder and forecaster
  10. Experimentation
  11. Benchmark of proposed approaches
  12. Long term prediction
  13. Perspectives
  14. Conclusions

Figures (7)

  • Figure 1: MOCAT-ML is designed to propagate phase space density distributions up to epoch $t + n + m$, given a preliminary sequence of distributions up to $t + n$, which are computed by a reliable and accurate method such as MOCAT-MC.
  • Figure 2: Encoder-decoder architecture to lower the dimensionality of the input space before applying a Deep Neural Network (DNN) suitable for time series forecasting.
  • Figure 3: Overview of the encoder-decoder ConvGRU neural network. Encoder blocks (three in this example) downsample and encode temporal features, which are then upsampled and decoded by decoder blocks (also three) to predict the output sequence. Temporal coherence is maintained by ConvGRU cells, while spatial resolution is restored via PixelShuffle upsampling.
  • Figure 4: Comparison of the predictions (along with last input of the lookback sequence) of the sample with worst validation loss among the different proposed approaches
  • Figure 5: Comparative evolution of loss functions for overlapping (2 weeks per step) and non-overlapping (2 months per step) forecasting strategies: full period and first year analysis.
  • ...and 2 more figures