Machine Learning Maps Hidden Orbit Structures in Four-Dimensional Space

For more than a century, the three-body problem has stood as one of the most famously intractable puzzles in celestial mechanics. When a spacecraft drifts through the gravitational landscape of two large bodies, such as the Earth and the Moon, its path can bend, loop, and scatter in ways that defy simple prediction. Now, a team of researchers at the Air Force Institute of Technology has unveiled a machine learning pipeline that can automatically classify the hidden architecture of these trajectories, potentially transforming how mission designers chart routes through cislunar space and beyond. The study, published in Astrophysics and Space Science, applies unsupervised clustering to four-dimensional Poincaré maps in the circular restricted three-body problem, or CR3BP, and demonstrates that algorithms can recover meaningful dynamical structures from data that would overwhelm even the most patient human analyst.

The CR3BP reduces the gravitational dance of three bodies to its essential form: an infinitesimally small mass, such as a spacecraft, moving under the influence of two primaries that orbit their shared center of mass in perfect circles. Within this simplified but still chaotic system, trajectories fall into three fundamental categories. Periodic orbits repeat themselves exactly, tracing closed loops through phase space. Quasi-periodic orbits wander across the surface of an invisible torus, never quite repeating but never straying far from home. Chaotic trajectories, by contrast, diverge unpredictably from their initial conditions, sensitive to the smallest perturbation. Distinguishing among these behaviors is the central task of trajectory design, because a mission planner must know whether a candidate path will remain stable or spiral into chaos.

The traditional tool for this task is the Poincaré map, a technique dating back to Henri Poincaré’s foundational work in 1890. Rather than tracking a trajectory continuously, the map records only the points where the path pierces a chosen surface in phase space, converting a continuous curve into a discrete scatter of crossings. In the planar version of the problem, this produces a two-dimensional plot in which periodic orbits appear as fixed points, quasi-periodic orbits form distinctive chains of islands, and chaotic motion scatters into structureless dust. Mission designers have long relied on visual inspection of these maps to identify promising orbits, a process that works well enough in two dimensions but collapses entirely when the problem extends into three spatial dimensions.

Extending the Poincaré map to four dimensions is necessary because real spacecraft do not confine themselves to planes. To capture out-of-plane motion, researchers use a space-plus-color representation: two spatial coordinates occupy the horizontal axes, the vertical coordinate is plotted on a third axis, and the out-of-plane velocity is encoded as color. The result is a rich but visually cluttered dataset in which the familiar island chains of two-dimensional maps vanish, replaced by three-dimensional structures with no established taxonomy. Previous efforts to sift through these maps relied on filtering by the Fast Lyapunov Indicator, a numerical measure of chaos, but this remained a manual, time-intensive operation that yielded only preliminary catalogs of structures. With each map containing thousands to millions of data points, the need for automation became unmistakable.

The new pipeline, developed by Kevin M. Trigg, Daniel J. Broyles, Robert A. Bettinger, and Tyler J. Kapolka, addresses this challenge through a carefully engineered sequence of computational steps. The researchers simulated 10,000 seed trajectories for each of 11 different Jacobi constants, a parameter that acts like an energy level and determines which regions of phase space a spacecraft can reach. Each trajectory was propagated for 1,000 time units using high-precision numerical integration, with early termination for paths that collided with a primary body or escaped the system. The team then faced a fundamental data problem: trajectories produce variable numbers of Poincaré crossings, and clustering algorithms require inputs of fixed length. Their solution was a feature engineering scheme that compresses each trajectory into a 23-dimensional numerical fingerprint.

That fingerprint draws on three complementary strategies. Statistical and dynamical descriptors capture the bounding box of each trajectory’s crossings across position and velocity coordinates, along with the Fast Lyapunov Indicator, which quantifies sensitivity to initial conditions and separates regular from chaotic behavior. Geometric descriptors treat each trajectory’s crossings as a miniature dataset, applying clustering within the trajectory itself to measure how cohesive or fragmented its structure is; a well-formed invariant torus yields dense, orderly crossings, while chaos produces scattered noise. Finally, frequency-domain analysis applies a Fast Fourier Transform to the ordered sequence of crossings, extracting the three strongest oscillatory modes and their amplitudes as signatures of recurrence. Periodic and quasi-periodic trajectories concentrate their spectral energy in sharp peaks, whereas chaotic trajectories spread it broadly.

With feature vectors in hand, the pipeline applies Principal Component Analysis to compress the representation, retaining components that explain at least 95 percent of the variance and typically reducing 23 features to 11. The reduced vectors then feed into HDBSCAN, a hierarchical density-based clustering algorithm chosen for its ability to handle clusters of varying density and shape without requiring the number of clusters to be specified in advance. This flexibility matters because Poincaré maps contain structures with irregular boundaries and non-uniform density, conditions that defeat simpler methods like DBSCAN, which relies on a single global density threshold. HDBSCAN also explicitly labels low-density trajectories as noise, providing a natural mechanism for isolating chaotic orbits that belong to no coherent structure. The researchers compared HDBSCAN against agglomerative clustering, spectral clustering, affinity propagation, and Gaussian mixture models, finding that while competitors achieved slightly higher Silhouette scores, HDBSCAN consistently produced lower Davies-Bouldin scores, better structural similarity, and the crucial advantage of automatic noise detection.

The analysis revealed eight distinct dynamical structures in the four-dimensional maps, which the authors named descriptively in the absence of any formal taxonomy: Figure-Eight, Tube, Scorpion Tail with Dots, Scorpion Tail without Dots, Pillar and Shield, Pillar and Shapes, Two Planes, and Three Pillars. These geometries bear no resemblance to the island chains of planar maps, underscoring how radically the topology changes when out-of-plane motion is included. The Figure-Eight structures, associated with quasi-periodic motion around Lagrange points, dominated the regular regions of the maps. An ablation study confirmed that every feature group contributed essential information: removing the Fast Lyapunov Indicator degraded the separation between regular and chaotic regions, while removing statistical features collapsed intra-cluster cohesion. When the tuned clustering configuration was applied unchanged to maps at other Jacobi constants, it recovered consistent structures with noise proportions holding steady between roughly 20 and 30 percent, demonstrating genuine robustness across the energy landscape.

Perhaps the most ambitious component of the work is its approach to orbit continuation. In classical astrodynamics, tracing how an orbit family evolves requires differential correction and continuation methods that follow a trajectory as parameters change. The researchers instead built a data-driven approximation: they treated each Jacobi constant level as a layer in a directed graph, connected each trajectory to its three nearest neighbors in feature space within the adjacent layer, and used depth-first search to find chains spanning the entire range from C equals 3.18 down to 2.68. Chains were ranked by average feature distance, with the smoothest paths representing candidate continuations of dynamical families. The results confirmed that the learned feature representations preserve meaningful dynamical similarity across energy levels, with Figure-Eight structures persisting coherently across maps. Yet the method also exposed its own limitation: only 24 unique structures appeared among the top 1,000 chains, revealing a strong bias toward the dominant, tightly clustered geometries and a need for additional constraints to encourage exploration of rarer structures.

The implications extend beyond the Earth-Moon system. The authors emphasize that the methodology applies to any multi-body gravitational environment, and that four-dimensional Poincaré maps represent a growing frontier in astrodynamics research. Future work includes cataloging how these structures evolve with mass parameter and energy, comparing the discovered orbits against known periodic orbit families and invariant tori generated by traditional continuation methods, and developing algorithms to locate fixed points in maps where, unlike the planar case, they are not intuitively positioned at the centers of island chains. The pipeline also contributes to a broader scientific goal: automated detection of invariant structures in high-dimensional Hamiltonian systems, a challenge that reaches into plasma physics, celestial mechanics, and accelerator design. For mission planners navigating the increasingly crowded cislunar arena, where spacecraft such as those supporting lunar exploration must exploit subtle gravitational structures to conserve fuel, a tool that converts millions of trajectory crossings into organized, labeled dynamical families could shorten the path from concept to flight-ready trajectory design.

Subject of Research: Unsupervised machine learning for discovering dynamical structures in 4D Poincaré maps of the circular restricted three-body problem

Article Title: Application of machine learning to discover dynamical structures in 4D Poincaré maps in the circular restricted three-body problem

Article References: Trigg, K. M., Broyles, D. J., Bettinger, R. A., & Kapolka, T. J. (2026). Application of machine learning to discover dynamical structures in 4D Poincaré maps in the circular restricted three-body problem. Astrophysics and Space Science, 371(9), Article 107. https://doi.org/10.1007/s10509-026-04638-5

Image Credits: AI Generated

DOI: 10.1007/s10509-026-04638-5

Keywords: Poincaré map, circular restricted three-body problem, machine learning, HDBSCAN, quasi-periodic orbits, chaos, Fast Lyapunov Indicator, principal component analysis, astrodynamics, Earth-Moon system, trajectory design, clustering