<p>Initial orbit determination (IOD) is an increasingly relevant problem in the cislunar regime, where chaotic dynamics degrade the performance of classical IOD approaches. In this investigation, a framework for an end-to-end cislunar IOD process is presented that incorporates angles-only observations that simulate a chance detection scenario. Enlisting machine learning techniques to assist the complicated cislunar IOD process, this work presents the Machine Classifier for Cislunar Orbit Determination (MCCLOD) model that employs a neural network for infusing information about known multi-body dynamics structure about the circular restricted three-body problem (CRTBP) into the initial state estimation process. The novel MCCLOD IOD process is compared directly with a classical two-body IOD approach (Gooding) for Earth–Moon <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40295_2025_525_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$L_{1}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>L</mi> <mn>1</mn> </msub> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="40295_2025_525_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$L_{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>L</mi> <mn>2</mn> </msub> </math></EquationSource> </InlineEquation> halo orbit examples in a high-fidelity dynamics environment. Two neural network (NN) models are implemented with regression and orbit classification to identify a 6D state estimate given angles-only measurements. The resulting simulations indicate drastic improvement in both accuracy (MCCLOD demonstrates at best two orders of magnitude improvement in positional error performance) and batched least-squares convergence consistency. Although a “classification with regression” NN degrades overall MCCLOD IOD performance, simulations indicate that a hybrid classification NN followed by a regression NN framework yields low position error in tested cislunar IOD problems.</p>

An Adaptive Approach to Cislunar Initial Orbit Determination Using Machine Learning

期刊论文
发表日期：

2025年9月4日

查看全文

Juan Ojeda Romero,
Wayne Schlei,
Gene Whipps,
Nick LaFarge,
Gunner Fritsch,
Sean Phillips

摘要

Initial orbit determination (IOD) is an increasingly relevant problem in the cislunar regime, where chaotic dynamics degrade the performance of classical IOD approaches. In this investigation, a framework for an end-to-end cislunar IOD process is presented that incorporates angles-only observations that simulate a chance detection scenario. Enlisting machine learning techniques to assist the complicated cislunar IOD process, this work presents the Machine Classifier for Cislunar Orbit Determination (MCCLOD) model that employs a neural network for infusing information about known multi-body dynamics structure about the circular restricted three-body problem (CRTBP) into the initial state estimation process. The novel MCCLOD IOD process is compared directly with a classical two-body IOD approach (Gooding) for Earth–Moon $L_{1}$ $L_{1}$ and $L_{2}$ $L_{2}$ halo orbit examples in a high-fidelity dynamics environment. Two neural network (NN) models are implemented with regression and orbit classification to identify a 6D state estimate given angles-only measurements. The resulting simulations indicate drastic improvement in both accuracy (MCCLOD demonstrates at best two orders of magnitude improvement in positional error performance) and batched least-squares convergence consistency. Although a “classification with regression” NN degrades overall MCCLOD IOD performance, simulations indicate that a hybrid classification NN followed by a regression NN framework yields low position error in tested cislunar IOD problems.