Clustering Astronomical Orbital Synthetic Data Using Advanced Feature Extraction and Dimensionality Reduction Techniques
| Authors | Eraldo Pereira Marinho et al. |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2603.13177 |
| Download | |
| Categories | cs.AI |
Abstract
The dynamics of Saturn's satellite system offer a rich framework for studying orbital stability and resonance interactions. Traditional methods for analysing such systems, including Fourier analysis and stability metrics, struggle with the scale and complexity of modern datasets. This study introduces a machine learning-based pipeline for clustering approximately 22,300 simulated satellite orbits, addressing these challenges with advanced feature extraction and dimensionality reduction techniques. The key to this approach is using MiniRocket, which efficiently transforms 400 timesteps into a 9,996-dimensional feature space, capturing intricate temporal patterns. Additional automated feature extraction and dimensionality reduction techniques refine the data, enabling robust clustering analysis. This pipeline reveals stability regions, resonance structures, and other key behaviours in Saturn's satellite system, providing new insights into their long-term dynamical evolution. By integrating computational tools with traditional celestial mechanics techniques, this study offers a scalable and interpretable methodology for analysing large-scale orbital datasets and advancing the exploration of planetary dynamics.
Engineering Breakdown
Plain English
This paper tackles the problem of analyzing orbital dynamics in Saturn's moon system by building a machine learning pipeline to cluster ~22,300 simulated satellite orbits. The authors use MiniRocket, a time-series feature extractor, to convert 400-timestep orbital trajectories into a 9,996-dimensional feature space, then apply automated feature extraction and dimensionality reduction to enable robust clustering. Traditional methods like Fourier analysis fail at this scale and complexity, but the ML approach successfully identifies orbital patterns and resonance interactions that would be invisible to classical stability metrics.
Core Technical Contribution
The core novelty is applying MiniRocket—a rocket-based time-series classifier known for speed and accuracy—to orbital dynamics problems where it hasn't been used before. Rather than hand-engineering Fourier features or stability indices, the authors let MiniRocket automatically discover the feature patterns in raw 400-timestep orbital sequences, outputting a 9,996-dimensional representation that captures nonlinear temporal structure. The pipeline then chains automated feature engineering and dimensionality reduction to make the high-dimensional features tractable for clustering, turning a previously intractable problem (analyzing 22K+ orbits with classical methods) into a solvable one.
How It Works
The input is 22,300 simulated satellite orbits, each represented as a time series of 400 timesteps (likely position/velocity states). These trajectories are fed into MiniRocket, which applies multiple random convolutional kernels at different scales and dilations, generating 9,996 summary features per orbit without any training required—this is why it's so efficient. Next, automated feature extraction (likely filter-based or statistical methods) selects the most informative features from that 9,996-dimensional space, reducing noise and redundancy. Finally, dimensionality reduction (e.g., PCA, t-SNE, or UMAP) compresses these features to 2D or 3D for clustering algorithms (k-means, DBSCAN, etc.) to identify orbital families and resonance groups. The output is a set of clusters where orbits within each cluster share similar dynamical behavior, enabling scientists to understand which orbital configurations are stable and how they interact.
Production Impact
For engineers building physics-informed ML systems, this approach offers a template for analyzing high-dimensional temporal data without domain-specific feature engineering—you can process 22K+ simulation trajectories end-to-end without manual Fourier decomposition or hand-coded stability metrics. In a real pipeline, you'd replace custom feature extraction scripts with MiniRocket, reducing engineering effort and improving reproducibility. The trade-off is computational: MiniRocket is fast (~milliseconds per orbit) but generates a very high-dimensional intermediate representation (9,996 features); you'd need to budget for the dimensionality reduction step and ensure your clustering algorithm can handle it. For real-time orbital prediction systems, MiniRocket's inference speed is a major advantage—you can extract features from new trajectories on-the-fly—but the 22K training set size is a lower bound; production deployments handling millions of orbits would need to verify scaling behavior.
Limitations and When Not to Use This
The paper does not address how to choose hyperparameters for MiniRocket (kernel count, dilation ranges, etc.) or justify the 9,996-dimensional choice; in production, you'd need ablation studies to understand sensitivity. The pipeline is validated only on simulated Saturn orbits, so generalization to real observational data (with noise, missing values, and perturbations from moons and rings) is untested. The paper likely doesn't cover validation metrics beyond clustering cohesion—you'd want to verify that discovered clusters actually correspond to physically meaningful orbital families (resonances, families, chaos zones) using domain expert review. Additionally, the approach assumes orbital states can be captured in 400 timesteps; shorter or longer observation windows, or orbits with multi-timescale behavior, may require retuning and could break the pipeline.
Research Context
This work builds on a wave of recent success with MiniRocket and rocket-based methods in time-series classification, which have outperformed traditional feature engineering on UCI and other benchmarks. It extends MiniRocket beyond its original domain (UCR time-series classification) into orbital mechanics and unsupervised learning (clustering), showing the versatility of learned convolutional features for scientific simulation data. The paper sits at the intersection of physics-informed machine learning and time-series analysis, opening a research direction where classical chaos/stability tools (Lyapunov exponents, KAM theory) are augmented or replaced by unsupervised ML for fast, scalable orbit classification on massive simulation datasets.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::