Long-Term Dynamical Evolution and Ejection of Near-Earth Asteroids

1. Problem Statement

Predicting the long-term dynamical fate of Near-Earth Asteroids (NEAs)—specifically whether they will remain gravitationally bound to the Solar System or be ejected over million-year (Myr) timescales—is critical for planetary defense and understanding Solar System evolution.

• The Challenge: High-precision $N$-body numerical integrations (using methods like Bulirsch-Stoer) are computationally expensive and do not scale well to large asteroid populations.
• The Nature of the Problem: NEA dynamics are highly chaotic with short Lyapunov times ($\sim 100\text{--}200$ years), making deterministic long-term prediction impossible. Instead, outcomes must be treated statistically.
• The Gap: While Machine Learning (ML) has been used for asteroid classification, its application to predicting long-term ejection outcomes using both static orbital elements and dynamic trajectory signatures remains limited.

2. Methodology

The authors propose a dual-framework approach comparing traditional ML classifiers with Deep Learning (DL) models to predict ejection within a 1 Myr window.

A. Data Generation and Preprocessing

• Numerical Integration: Used the MERCURY package with the high-accuracy Bulirsch-Stoer (BS) integrator.
  • Model: Sun + 8 planets + Pluto as massive bodies; asteroids as massless test particles.
  • Settings: 1 Myr integration, 8-day timestep, $10^{-12}$ tolerance. Non-gravitational forces (Yarkovsky-like) were included.
  • Dataset: 35,455 NEAs from the JPL Small-Body Database.
  • Uncertainty Handling: Generated Monte Carlo ensembles (50 clones per object) to test robustness against observational uncertainties.
• Labeling: An asteroid is labeled "Ejected" if it exceeds 100 AU or attains hyperbolic escape velocity within 1 Myr; otherwise, it is "Non-Ejected."
• Class Imbalance: Only ~15% of asteroids were ejected. The dataset was augmented via controlled perturbations of initial orbital elements and temporal segmentation (splitting 1 Myr integrations into 0.2 Myr windows) to balance the training set.

B. Feature Engineering

Two distinct feature sets were used:

1. Static Features (for ML): Initial orbital elements at the epoch (semi-major axis $a$, eccentricity $e$, inclination $i$, argument of perihelion, etc.).
2. Dynamic Features (for DL): Recurrence Plots (RPs) derived from short-term (0.2 Myr) time series of $a$ and $e$.
   • RPs transform 1D time series into 2D binary images (150 $\times$ 150 pixels) encoding temporal recurrence, periodicity, and chaotic divergence.
   • These images serve as inputs to a Convolutional Neural Network (CNN).

C. Models Evaluated

• Traditional ML (7 Models): K-Nearest Neighbors (KNN), Decision Trees, Random Forests (RF), ExtraTrees, Multilayer Perceptrons (MLP), AdaBoost, and Gradient Boosting (GB).
• Deep Learning: A custom Convolutional Neural Network (CNN) with three convolutional layers (32, 64, 128 filters), max-pooling, and fully connected layers, trained on the RP images.

3. Key Contributions

1. Hybrid Framework: First study to systematically compare static orbital element classifiers against CNNs trained on recurrence plot representations of short-term orbital evolution for NEA ejection prediction.
2. Recurrence Plot Application: Successfully demonstrated that RPs derived from short (0.2 Myr) integrations can encode sufficient dynamical information to predict 1 Myr ejection outcomes, capturing chaotic signatures invisible to static analysis.
3. Backward Integration Analysis: Performed backward integrations to diagnose time-asymmetry in chaotic regions. Found partial overlap between forward- and reverse-ejected sets, identifying a subset of objects in highly unstable phase-space regions.
4. Scalable Proxy: Established that data-driven models can serve as efficient proxies for expensive $N$-body integrations, enabling the prioritization of asteroids for detailed study.

4. Key Results

Performance Metrics

• Traditional ML: Tree-based ensemble models (Random Forest, AdaBoost, Gradient Boosting) performed best.
  • Accuracy: $\sim 86\%$
  • Recall: $\sim 91\%$ (crucial for identifying hazardous ejected objects)
  • F1 Score: $\sim 0.87$
  • Feature Importance: Semi-major axis ($\sim 72\%$) and eccentricity ($\sim 13\%$) were the dominant predictors, accounting for ~85% of the decision power.
• Deep Learning (CNN):
  • Accuracy: $\sim 87\%$ (slightly higher than ML)
  • AUC-ROC: 0.94
  • Recall: $\sim 88\%$
  • The CNN successfully learned temporal signatures of chaos from the RP images, achieving performance comparable to or slightly better than the best ML models.

Dynamical Insights

• Ejection Patterns: Ejected asteroids cluster in regions of high semi-major axis ($1.5\text{--}3.5$ AU) and high eccentricity ($0.4\text{--}0.8$), primarily belonging to the Apollo and Amor groups.
• Temporal Evolution: Ejection rates peak early (0.2–0.4 Myr) and decline, indicating that the most unstable objects are removed quickly.
• Time Asymmetry: Backward integrations revealed that while ~15% of objects are ejected in both directions (highly unstable), many objects are ejected only in the forward direction (recently destabilized) or only backward (recently captured).
• Model Generalization:
  • ML models (relying on static initial conditions) maintained high performance on backward-integrated data.
  • CNN performance dropped in precision on backward data due to distribution shifts in the RP features, highlighting the sensitivity of DL to input distribution changes.

5. Significance and Implications

• Computational Efficiency: The study validates that tree-ensemble models trained on initial ephemerides offer a highly efficient, interpretable, and low-cost method for screening large asteroid catalogs.
• Complementary Approaches: While ML is faster, the CNN approach captures deeper temporal dynamics. The authors propose a hybrid strategy: use ML for rapid, large-scale triage and reserve computationally intensive trajectory-based (CNN/RP) analysis for ambiguous or high-value targets.
• Planetary Defense: These methods provide scalable tools to prioritize asteroids for detailed long-term dynamical studies, improving the efficiency of impact hazard assessment without requiring exhaustive $N$-body simulations for every object.
• Limitations: The models are statistical proxies; they cannot predict exact ephemerides. They may miss "slow-burn" instabilities that develop after the 0.2 Myr training window, and results are subject to observational uncertainty and numerical modeling choices (e.g., collision handling).

In conclusion, the paper demonstrates that machine learning, particularly ensemble tree methods and CNNs trained on recurrence plots, provides a robust, scalable framework for predicting the long-term dynamical fate of Near-Earth Asteroids, effectively bridging the gap between computational cost and dynamical insight.