AI-driven collision risk prediction and debris optimization in low earth orbit Authors: Emad Alkhazraji, Mohamed E. Yahia, R. E. Tolba, N. S. Alharthi Published: 2026 ABSTRACT Debris congestion and collision in low Earth orbit have become a major point of concern due to the rapid expansion of activities in that region. In this work, an AI-driven framework is developed and deployed to predict, analyze, and optimize collision risk across the altitude inclination space. Local debris object population, mass density distribution, and conjunction derived risk are computed. Then, an AI surrogate model, viz., gradient-boosted decision trees, was trained to predict collision conjunction relative probabilities and risk. Thereafter, data analysis and explain ability techniques are utilized to uncover underlying risk drivers and reveal how altitude, inclination, mass loading, and crowding interact across the orbital space. The analyses showed how the AI surrogate model reliably captures underlying nonlinear risk behavior and interactions. More importantly, it suggested that meaningful capacity increases are achievable with only a moderate rise in predicted risk. Finally, the AI surrogate model was employed in a multi-objective evolutionary optimization to quantify the trade off between minimizing collision risk and maximizing usable orbital capacity, revealing a set of optimal, feasible, and well-balanced points of operation in this regime. Ultimately, the proposed pipeline offers a computationally efficient pathway for prediction and interpreting orbital safety, guiding constellation deployment, and supporting sustainable space traffic management strategies. Introduction The proliferation of human-made objects in the low Earth orbit (LEO) has led to a significant increase in space debris, posing substantial risks to operational spacecraft. Space debris encompasses non-functional objects that range from defunct satellites and spent rocket stages to smaller fragments resulting from collisions or disintegration events1–3. In 2022, over 70% of all objects in the LEO were classified as debris, underscoring its critical density in this orbital region2,4. The primary concern associated with this accumulation is the potential for collisions, which can generate even more debris and potentially trigger Kessler syndrome, a cascade effect whereby collisions with debris create more debris, leading to further collisions and rendering certain orbital regions unusable2. This is particularly relevant in the LEO, which is heavily utilized for various space activities, including the deployment of large satellite constellations5,6. No doubt that the explosive increase in man-made objects in low Earth orbit (LEO) has led to a substantial increase in space debris. This poses a significant risk to operational spacecraft, satellite missions, and the long-term sustainability of space activities1,2,7,8. The international community is also grappling with the legal and governance challenges space debris poses. Increasing congestion in near-Earth orbits, driven by the rapid growth of satellites and debris, directly threatens the sustainable use of space9. Legal frameworks such as the 1967 Space Treaty are being examined to understand their applicability to managing and mitigating satellite debris in the LEO8. There is a recognized urgency for addressing space debris, with international bodies like the United Nations Committee on the Peaceful Uses of Outer Space taking steps to inform states about the importance of preventing debris generation10. Various approaches have been taken to address the challenge of space debris, including detection and tracking. Traditional ground-based equipment faces limitations in observing fast-moving LEO debris due to atmospheric interference and restricted fields of view11. To overcome these challenges, new technologies, such as optical space-based surveillance networks for tracking LEO debris, are being explored11. Debris removal is also crucial. One promising area involves using miniaturized satellites, such as CubeSats, with robotic arms to deorbit larger debris objects4. This approach is economically viable and offers shorter design-to-deployment times compared to larger missions4. The guidance for such missions is being optimized through convex-optimization-based model predictive control algorithms, which account for factors such as J2 perturbation, drag, and eclipses12. Moreover, risk management and control remain paramount pre- and during missions13,14. Artificial intelligence (AI) and machine learning (ML) techniques are being developed to improve the accuracy of orbital lifetime estimation for LEO satellites, which is required to comply with debris mitigation guidelines, such as the 25-year deorbit rule established by the Inter-Agency Space Debris Coordination Committee (IADC)15. Traditional numerical orbit propagation methods, while accurate with high-precision force models, can be computationally intensive. ML approaches are faster and comparably accurate, thus enabling a global assessment of guideline compliance15. In collision avoidance, AI-driven optimization frameworks are essential for managing the inherent uncertainty in collision risk assessment16. Space debris poses a significant threat, increasing the likelihood of mission failure or catastrophic damage. AI assists in optimizing collision avoidance maneuvers by accounting for uncertainties in factors such as conjunction probabilities and orbital perturbations16. For the detection and tracking of space debris, AI has been integrated into various advanced systems, including space-based surveillance networks11, radar systems17, LiDAR sensors18, density model optimization19, mission planning for active removal20, and rendezvous trajectory generation21. In this work, we employ an ML model and an AI surrogate to predict the conjunction probability of and risk from LEO debris. The surrogate model acts as a fast, efficient way to predict newly detected debris objects. Furthermore, the model is then utilized to further analyze different parameters that affect the conjunction factors using AI analysis techniques, namely principal component analysis (PCA) and SHapley Additive exPlanations (SHAP). Lastly, enabled by its fast and efficient nature, the AI surrogate model is then utilized to perform a genetic algorithm to optimize different conditions and parameters. The contribution of this work is not the introduction of new individual methodologies, but rather their structured integration into a unified workflow that links prediction, interpretation, and optimization within a single consistent framework. This integration enables a more comprehensive analysis of orbital risk by connecting model outputs with interpretable feature contributions and decision-oriented trade-off exploration. While prior studies have addressed collision risk estimation, data-driven modeling, or optimization independently, the proposed approach provides a cohesive pipeline that enables these components to be applied jointly, allowing for improved interpretability and decision support. Methodology In this work, we utilized the LEO−Catalog dataset cordially provided by LEOLABS22. The dataset catalogued 25,000+ debris objects in LEO, categorized into fragments, payloads, and rocket bodies, detailing each object’s launch date, country of origin, apogee, perigee, and inclination. The catalog includes records spanning from 1959 to 2025, reflecting the historical evolution of the orbital environment. However, the present analysis is conducted using a static snapshot of the dataset at the time of acquisition to ensure consistency in model training and evaluation. In other words, temporal dynamics are not explicitly modeled in this work. This approach allows the study to focus on spatial and population-based risk patterns rather than temporal forecasting. Missing values were handled using a combination of filtering and simple imputation strategies. Non-critical missing entries were replaced using default or zero-filled values where appropriate, while ensuring that all key variables required for risk estimation were retained and validated. Data preparation and processing It is a common practice to divide the modelling space into bins. In our case, we adopted the NASA Orbital Debris Engineering Model (ORDEM) in LEO standards and opted for 45 bins of altitude (40 km) and inclination (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4^\circ$$\end{document})23. The altitude was estimated as the average between the apogee and perigee. The goal of this study is to analyze and optimize-engineer the collision probability and risk conjunctions in LEO. Assuming a Gaussian distribution of position uncertainty, the probability of collision (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_c$$\end{document}) is approximated as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{c} = \frac{1}{{2\pi \sqrt {|{\mathbf{C}}|} }}\int_{{V_{c} }} {\exp } \left( { - \frac{1}{2}{\mathbf{r}}^{{ \top }} {\mathbf{C}}^{{ - 1}} {\mathbf{r}}} \right)d{\mathbf{r}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V_c$$\end{document} is the collision volume, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{C}$$\end{document} is the combined covariance matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{r}$$\end{document} is the relative position vector. However, such an integral has no closed form, and it is common to utilize a spatial risk proxy instead of the absolute \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_c$$\end{document}24. In classical formulations, collision probability is derived from an integral over relative motion, object density, and encounter cross-section, accounting for spatial distribution and velocity dispersion. In this work, a simplified proxy is adopted by aggregating objects into discrete altitude–inclination bins and assuming locally uniform distributions within each bin. Under these assumptions, the collision likelihood can be approximated as proportional to the local object density and population. This simplification introduces deviations from the full integral formulation due to several assumptions, including spatial homogeneity within bins, neglect of detailed relative velocity distributions, and discretization of the orbital domain. As a result, the proposed metric does not represent an absolute collision probability, but rather a normalized indicator of relative collision likelihood. Despite these simplifications, the proxy preserves the relative variation of collision risk across different regions of the orbital environment, which is the primary objective of this study. Therefore, it provides a computationally efficient and physically informed metric suitable for comparative analysis and optimization. As such, the probability of a conjunction, as a risk index, can then be approximated as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P_{\text {conj}} \propto \rho \, v_r\ A_{\text {eff}} \ \Delta t \end{aligned}$$\end{document}where ρ is the local object density (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {objects/km}^{3}$$\end{document}), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{rel}$$\end{document} is the typical relative speed (~10 km/s in LEO), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{eff}$$\end{document} is the effective collision cross-section (~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10\ m^2$$\end{document}), while Δ t is the time window (1 day in this study). It is important to note that this formulation does not represent a fully calibrated physical collision probability. Instead, it provides a normalized proxy that captures relative encounter likelihood based on local population and density characteristics. The term “collision probability proxy” is therefore used throughout this work to emphasize its comparative nature. Lastly, the risk (R) is estimated as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R = P_{\text {conj}} \times M_{ij} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_{ij}$$\end{document} is the combined mass between two objects i and j. The resulting risk metric should be interpreted as a relative indicator that combines encounter likelihood with potential impact severity through aggregated mass. It does not represent an absolute probability or expected collision frequency, but rather enables consistent comparison across orbital bins. To facilitate carrying out Equation 3, the mass of catalogued objects was approximated according to the European Space Agency (ESA) MASTER-8 (Meteoroid and Space debris Terrestrial Environment Reference)25. That said, the objective of this formulation is not to predict absolute collision rates, but to construct a physically informed and computationally efficient surrogate metric that preserves the relative structure of collision risk across the orbital environment. As such, the right semi-circle in Fig. 1 (a) shows object count per altitude and inclination bins, whereas the left semi-circle shows estimated total mass per altitude and inclination bins. The right semi-circle in Fig. 1 (b) depicts the approximated probability proxy of a conjunction according to Equation 2, whereas the left semi-circle illustrates the risk according to Equation 3 per altitude and inclination bins. Both Figs. 1 (a) and (b) show the values on a logarithmic scale to display both large and small values. At first glance, said figures demonstrate that the majority of the presence of LEO debris objects is concentrated in inclinations between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$40^\circ$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100^\circ$$\end{document}, and the highest presence is located at an altitude of 500 km. In terms of Risk, spikes occur at sun-synchronous orbits (~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$97^\circ$$\end{document}), International Space Station-like orbits (~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$51^\circ$$\end{document}), and Molniya-like orbits (~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$74^\circ$$\end{document}).Fig. 1(a) Object count (right semi-circle) and total mass (left semi-circle) per altitude and inclination bins. (b) Estimated collision probability proxy (right semi-circle) and risk (left semi-circle) per altitude and inclination bins. (a) Object count (right semi-circle) and total mass (left semi-circle) per altitude and inclination bins. (b) Estimated collision probability proxy (right semi-circle) and risk (left semi-circle) per altitude and inclination bins. Surrogate modeling The next step in our analysis is to build and train surrogate models that predict the collision probability proxy and relative collision risk at each altitude and inclination bin, based on the variables (features) discussed in subsection 2.1, namely altitude, inclination, object density, count, and combined mass. In particular, the XGBoost (eXtreme Gradient Boosting) machine learning model was employed. The benefit of using the surrogate model is twofold: (1) Pose as a fast and efficient surrogate (replacement) model that can predict the conjunction probability proxy and risk of any newly detected debris object. (2) Open the door to further analysis of the dependencies and correlation between the given features and conjunction probability proxy and relative risk (as discussed in the next section). XGBoost is a highly efficient, machine learning model for supervised tasks like classification and regression, known for its speed, performance, and scalability, especially with large datasets like the one at hand. XGBoost was selected as the surrogate modeling approach due to its strong performance on structured tabular datasets and its ability to capture nonlinear relationships and feature interactions. Compared to traditional ensemble methods such as Random Forests, gradient boosting models typically provide improved predictive accuracy and better handling of complex dependencies. In addition, XGBoost offers efficient training and seamless integration with explainability frameworks such as SHAP, which are essential for interpreting model predictions in this study. While alternative models such as LightGBM and Random Forests are also widely used, the focus of this work is not on benchmarking different machine learning algorithms, but rather on demonstrating an integrated framework that combines prediction, interpretation, and optimization. Therefore, XGBoost is adopted as a representative high-performance model suitable for this application. Initially, the LEO Catalog dataset was divided as 80–20% between training and test datasets, respectively, for validation. Table 1 summarizes the hyperparameters adopted for the training process.Table 1Hyperparameters adopted for surrogate model training.HyperparameterValueNumber of estimators2000Maximum depth6Learning rate0.015Subsample ratio0.8Subsample ratio of columns0.9Regularization of the leaf weights3.0L2 regularization term on weights3.0Tree methodHistogram-based algorithmEvaluation metricsRoot Mean Square ErrorMean Absolute Error Table 1Hyperparameters adopted for surrogate model training.HyperparameterValueNumber of estimators2000Maximum depth6Learning rate0.015Subsample ratio0.8Subsample ratio of columns0.9Regularization of the leaf weights3.0L2 regularization term on weights3.0Tree methodHistogram-based algorithmEvaluation metricsRoot Mean Square ErrorMean Absolute Error Hyperparameters adopted for surrogate model training. The proposed framework is designed as a decision-support tool for analyzing orbital congestion and guiding satellite deployment strategies. The workflow consists of three main stages: data preprocessing, surrogate model prediction, and multi-objective optimization. Model training is performed offline and requires standard computational resources for gradient boosting models. Once trained, the surrogate model enables rapid inference, with prediction times on the order of milliseconds per sample. This allows efficient evaluation of large design spaces during optimization. The optimization process introduces additional computational cost, but remains tractable for moderate population sizes and generation counts. The framework is therefore suitable for scenario analysis and constellation design studies, where near real-time performance is sufficient. It is important to note that the proposed approach is not intended for real-time collision avoidance at the individual object level, but rather for higher-level risk assessment and strategic planning. Results and discussion In this section, we discuss the performance of the trained surrogate model and then use it in other AI modeling analysis methods, namely, PCA and SHAP. AI surrogate model performance Given the hyperparameters discussed in section 2, the surrogate AI model displayed a very high prediction accuracy on both training and validation datasets, signifying that no underfitting nor overfitting is exhibited. Figure 2 (a) shows the Root Mean Square Error (RMSE) after each iteration on both the training and validation sets, while Fig. 2 (b) shows the Mean Absolute Error (MAE) for both sets. After 500 iterations, the training process converged to a training RMSE of 0.084, an MAE of 0.037, with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} value of 0.996, whereas the validation RMSE was 0.132, the MAE was 0.055 with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} score of 0.964. These metrics demonstrate how strikingly well the trained AI model captures the underlying relation between the risk and adopted features, and prove that it can be used as a surrogate for the forthcoming analyses and optimization.Fig. 2AI training performance of the surrogate model in terms of (a) RMSE and (b) MAE. AI training performance of the surrogate model in terms of (a) RMSE and (b) MAE. Principal component analysis With the accurately trained AI model at hand, PCA was performed to uncover the underlying global structure of the feature space and unearth correlations and dominant modes of variation in the data on a macro population-level. PCA works by transforming the original correlated variables into a new set of orthogonal variables, called principal components (PCs). These PCs are linear combinations of the original features and are ordered such that the first few components retain most of the variation present in all of the original variables. Mathematically, the transformation is defined as26:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{Z} = \textbf{X} \times \textbf{V} \end{aligned}$$\end{document}where:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} \in \mathbb {R}^{n \times p}$$\end{document} is the zero-mean data matrix of n debris objects and p features,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V} \in \mathbb {R}^{p \times p}$$\end{document} is the matrix of eigenvectors (principal axes),\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Z} \in \mathbb {R}^{n \times p}$$\end{document} is the transformed data in the principal component space.Although PCA analysis provides a reduced-dimensional representation of feature variability and correlations in the dataset, it does not imply causal relationships between variables. The importance of each principal component is quantified by its associated eigenvalue \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _k$$\end{document}, representing the variance explained by the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{th}$$\end{document} PC:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Explained Variance Ratio}_k = \frac{\lambda _k}{\sum _{j=1}^{p} \lambda _j} \end{aligned}$$\end{document}Figure 3 (a) shows the correlation circle for the first (strongest) 2 PCs against the investigated parameters, while Fig. 3 (b) depicts the PC loading. PC1 represents how crowded and massive a region is, as it exhibits strong positive loadings from count, density, combined mass, and volume. Bins with high object count, high mass concentration, and small spatial volume cluster along this axis. This is a global indicator of collision likelihood and the severity of impact. In other words, PC1 is a congestion & mass driver, which is the largest variance component (capturing ~37% of the variance). On the other hand, dominated by inclination and altitude, PC2 captures the orbital architecture of LEO (~29% of variance), separating sun-synchronous orbits (~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$97^\circ$$\end{document}), ISS-like orbits (~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$51^\circ$$\end{document}), and mid-inclination clusters27.Fig. 3(a) PC correlation circle and (b) the PC feature loading first dominant 2 PCs. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} \in \mathbb {R}^{n \times p}$$\end{document} is the zero-mean data matrix of n debris objects and p features, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X} \in \mathbb {R}^{n \times p}$$\end{document} is the zero-mean data matrix of n debris objects and p features, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V} \in \mathbb {R}^{p \times p}$$\end{document} is the matrix of eigenvectors (principal axes), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V} \in \mathbb {R}^{p \times p}$$\end{document} is the matrix of eigenvectors (principal axes), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Z} \in \mathbb {R}^{n \times p}$$\end{document} is the transformed data in the principal component space. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Z} \in \mathbb {R}^{n \times p}$$\end{document} is the transformed data in the principal component space. (a) PC correlation circle and (b) the PC feature loading first dominant 2 PCs. SHAP interpretability SHAP is a technique that assigns a contribution value to each input feature for each prediction made by a surrogate model. It is based on cooperative game theory and provides local interpretability by aggregating feature contributions across all inputs. SHAP analysis is used here to interpret the behavior of the trained surrogate model by quantifying the contribution of each feature to the predicted output. It is important to note that these explanations reflect the internal structure of the model and do not, by themselves, establish causal relationships within the physical orbital environment. While SHAP does not provide causal inference, the identified feature contributions are consistent with established physical intuition, such as the influence of object density and population on collision likelihood. This agreement supports the credibility of the surrogate model while maintaining a clear distinction between correlation and causation. Figure 4 depicts the SHAP summary plot showing the effect each feature has on the predicted output (Risk). Features higher on the plot are generally more important, wherein each point represents a single instance’s SHAP value for a specific feature. The SHAP value for a given feature represents its contribution to the model prediction, and is computed as the average marginal contribution of the feature across all possible feature subsets27.Fig. 4SHAP summary plot quantifying the influence of a feature on the predicted risk across the entire design dataset. SHAP summary plot quantifying the influence of a feature on the predicted risk across the entire design dataset. While PCA revealed that the physical structure of the LEO is governed mainly by two axes: (1) congestion and mass concentration, and (2) orbital geometry defined by altitude and inclination, SHAP analysis complements this by quantifying how each feature affects the collision risk at the local (micro) level. In addition, the SHAP decision plots shown in Fig. 5 provide further insight into how each variable pushes the predicted risk for a sampled orbital bin. Specifically, Figs. 4 and 5 show that the combined mass, followed by count, then density, are the strongest drivers of risk. Particularly, congestion and mass concentration factors create distinct high-risk corridors in given altitude–inclination regimes. This corroborates the dominance of PC1 that was observed in Fig. 3. However, what SHAP analysis provides is a more detailed breakdown of how each feature ranks in affecting the overall output. Nonetheless, together with the PCA, SHAP provides a holistic explanation of risk formation, enabling interpretable AI-driven forecasting, hotspot detection, and targeted capacity-planning strategies for orbital debris management. Furthermore, underlying nonlinear effects are becoming clear; For instance, doubling mass more than doubles the risk, and the nonlinear factor of Density × Count interaction is strong as Risk surges when both are high. Moreover, the localized spread exhibited by altitude and inclination shows that Risk is not monotonic, but rather concentrated at specific operational regimes. This agrees with our intuition of the state of LEO and further corroborates the fact that the surrogate model accurately captures both global structure (PCA) and fine-scale effects (SHAP).Fig. 5Decision plot for sampled orbital bins, tracing how cumulative SHAP contributions steer each prediction from the prediction expected value. Decision plot for sampled orbital bins, tracing how cumulative SHAP contributions steer each prediction from the prediction expected value. Lastly, Fig. 6 depicts the SHAP dependence plots between each of the model’s input variables. They show how a single feature influences the model’s prediction across its full range, taking interactions with other features into account. In these plots, each point is colored by a second feature, which is the strongest interacting feature. They reveal whether the feature behaves differently for low vs high values of the interacting feature. It also shows any hidden nonlinear interactions (feature A affects the prediction only when feature B is in a certain range). For example, Inclination reduces risk at low densities but increases risk at high densities.Fig. 6SHAP dependence plots between the model’s variables taking interactions with other features into account. SHAP dependence plots between the model’s variables taking interactions with other features into account. Multi-objective optimization We conclude this study with formulating a multi-objective optimization problem and solving it using the Non-Dominated Sorting Genetic Algorithm II (NSGA-II). The trained surrogate model enables rapid prediction of collision risk, allowing it to reliably replace computationally heavy physics-based risk simulations during optimization iterations28,29. In particular, two conflicting objectives were adopted: (1) minimizing the collision risk, and (2) maximizing the usable orbital capacity, expressed as the available room for additional spacecraft with respect to a nominal capacity limit per orbital bin (in our case, 120% of the maximum observed count per bin). This capacity constraint is introduced as a relative measure of orbital crowding based on the observed population distribution. This reference value represents a controlled extrapolation beyond current occupancy levels, enabling the exploration of near-future congestion scenarios without extending into unrealistic regimes. This formulation does not imply a strict physical limit, but rather provides a normalized framework for comparing solutions within a bounded and physically interpretable design space. Similarly, the allowable variation range in object count is selected to prevent degenerate solutions while maintaining sufficient flexibility to explore meaningful trade-offs between population density and collision risk. Without constraints, the optimizer converges to trivial solutions in terms of unreasonable altitudes/inclinations, producing low-risk but unrealistic orbital configurations. Thus, we imposed physically motivated constraints that restrict the allowable deviation from the baseline to within a relative window of -50% to +200%. Table 2 shows the hyperparameters adopted in the NSGA-II algorithm.Table 2Hyperparameters adopted for the NSGA-II algorithm.HyperparameterValueNumber of generations80Population size120Crossover rate0.8Mutation rate0.2Selection methodCrowded Tournament Selection Table 2Hyperparameters adopted for the NSGA-II algorithm.HyperparameterValueNumber of generations80Population size120Crossover rate0.8Mutation rate0.2Selection methodCrowded Tournament Selection Hyperparameters adopted for the NSGA-II algorithm. Across the 80 generations, the algorithm evolves a diverse set of solutions that converge to a well-formed Pareto frontier (shown in Fig. 7). The frontier quantifies the inherent trade-off between collision safety and orbital capacity. Figure 7 shows that at the low-risk extreme of the front (left-hand side), risk values approach \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-10}$$\end{document}. However, these points correspond to very low usable capacity (handful per bin), illustrating the diminishing returns in reduced risk compared to the disproportionately large reductions required in population density. On the other hand, at the high-capacity end of the front (right-hand side), the predicted risk increases fairly to the order of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-8}$$\end{document}, revealing that substantial capacity gains can be achieved with only moderate increases in collision hazard. Between these extremes lies the middle ground that should be attractive to utilize. Overall, the resulting Pareto frontier provides a quantitative decision-support tool for satellite operators and regulators. It highlights specific orbital regimes where large increases in deployable capacity can be achieved with minimal added risk, as well as regions where even small reductions in risk require severe limitations on deployment density. By explicitly mapping the trade space, this surrogate-assisted NSGA-II framework enables informed planning of large-scale constellation deployments under realistic operational and safety constraints.Fig. 7Bi-objective optimization obtained with NSGA-II coupled to the AI surrogate model showing the Pareto frontier balancing between predicted risk and available capacity. Bi-objective optimization obtained with NSGA-II coupled to the AI surrogate model showing the Pareto frontier balancing between predicted risk and available capacity. Conclusion This work presented a holistic AI-powered framework that predicts, analyzes, and optimizes conjunction probability proxy and risk in Earth’s LEO. We demonstrated how AI can act as a robust pathway for both interpretability and computational scalability in space debris risk assessment. The surrogate model’s predictive fidelity is essential for large-scale case exploration, capacity planning, and optimization, where traditional physics-based models would be prohibitively slow or outright unfeasible. The subsequent data analysis methodologies we utilized exposed the dominant drivers shaping the LEO environment. The first component, responsible for ~37% of the total variance, reflected congestion (high population and high mass concentration), while the second component (~29% of the variance) captures orbital configuration. Later on, these findings were reinforced on a micro level by identifying combined mass, count, and density as the strongest local contributors to predicted risk. They also revealed important nonlinear interactions in play, such as the amplified risk when both density and mass rise simultaneously. This all culminated in the final stage, exploiting the practical and computational utility of the surrogate model multi-objective optimization. Competing goals were explored, namely minimizing conjunction risk and maximizing usable orbital capacity. The resulting Pareto frontier provides a rich map of feasible trade-offs. At the extreme low-risk end, the predicted risk values reached ~\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-10}$$\end{document} but required drastically reducing the population density. On the other end (high-capacity end), increasing the usable capacity towards the upper limit only increased the risk incrementally. Between these two extremes lies a sweet middle ground, or rather space, particularly suitable for sustainable constellation deployment. Overall, the proposed holistic framework offers a robust and interpretable decision tool, with the scalability and efficiency needed for real-time what-if analysis for regulators, constellation planners, and space traffic management entities.