A Symmetry-Reduced Differentiable Neural Information Field for Geometric Angle-Only Trackability Assessment Across Walker Low-Earth-Orbit Sensor Constellations Authors: Hengguo Zhang, Kebo Li, Yangang Liang, Yunxiao Lv Published: 2026 ABSTRACT Persistent angle-only tracking with Walker low-Earth-orbit constellations requires repeated geometric assessment, but optimal sensor-subset selection is computationally costly. To support efficient assessment, a symmetry-reduced differentiable neural information field is proposed to approximate the A-optimal potential of the best feasible sensor subset of a prescribed size. Its two-stage network replaces absolute time with bounded phases derived from axial and finite-permutation symmetries of a relative-periodic Walker constellation under J2. For a 25×25 constellation, symmetry reduction decreased the three-seed mean root-mean-square error from 7.664×10−3 to 1.154×10−3 relative to a matched two-body encoding; altitude-gradient correlation reached 0.9960. Finite-symmetry closure was verified for Delta, Star, and Rosette. Field accuracy remained within prescribed thresholds for 30 days under matched J2 dynamics. Without orbit maintenance, high-fidelity propagation yielded median and minimum sampled validity horizons of 1.50 and 1.25 days. On historical Iridium trajectories reconstructed from orbital records, three independently trained fields maintained the prescribed accuracy for at least 30 days, demonstrating long-horizon applicability to an operational Walker constellation. Neural triggering reduced penalized mean regret by 91.6% and candidate-pool searches by 93.4%. Full-grid neural inference achieved 17.5-fold CPU and 5567-fold GPU speedups over CPU candidate-pool evaluation. 1. Introduction Space-based optical sensors provide line-of-sight angular measurements for target surveillance, angles-only orbit determination, and cooperative spacecraft observation and navigation [1,2,3,4]. These applications require informative observation geometry as the relative positions of sensors and targets evolve. For a low-Earth-orbit (LEO) sensor constellation observing high-altitude moving targets, geometric assessment supports the identification of favorable observation regions and the reassignment of sensors along a target trajectory. These tasks require efficient assessment of how observation geometry varies across target locations and over time. Across these tasks, geometric visibility is necessary but not sufficient for informative angle-only measurements. Simultaneously visible satellites can provide nearly parallel lines of sight, leaving position information strongly anisotropic. Fisher information criteria quantify this effect in angle-only path planning, sensor selection, and cooperative orbital sensing [3,5,6]. Applying these criteria over large constellations and extended trajectories requires repeated assessment at many target locations and times. Direct evaluation at each query requires satellite positions, visibility screening, and the construction of Fisher information matrices (FIMs), followed by a search over admissible sensor subsets [7,8,9]. The resulting computational cost limits dense mapping and repeated tasking, while the individually evaluated samples do not directly provide a continuous field or its spatial gradients. Existing approaches address individual parts of this computational chain. Reinforcement learning, tree search, potential games, and mathematical scheduling support sensor selection and tasking, but still require repeated evaluation of geometry or uncertainty [10,11,12,13,14]. Constellation design studies similarly couple coverage with observation geometry [1], while optimization methods address constellation design decisions [15]. Fisher information fields organize sensing geometry as a reusable spatial representation for static visual localization [16], while neural surrogates reduce repeated computations in constellation analysis [17]. Coordinate-based neural fields with Fourier features provide continuous pointwise queries and spatial derivatives [18,19,20,21], and physics-informed models incorporate known structure into learned mappings [22,23]. Neural orbit models have focused on two-body solutions and orbital-element prediction [24,25]. For a neural information field of an orbiting sensor constellation, a further issue is how to represent temporal evolution while exploiting the relative periodicity of the sensing geometry. For a Walker constellation [26], the informative geometry changes jointly with target longitude, latitude, altitude, Earth rotation, orbital phase, and inter-plane phasing. Using absolute time as an additional network coordinate combines these motions in an unbounded variable and turns long-term evaluation into temporal extrapolation. The J2 perturbation adds nodal precession and short-period variation to the ideal Keplerian geometry [27,28]. These dynamics motivate a symmetry-based temporal representation, together with an assessment of how long the resulting field remains accurate when additional perturbations are present. This study develops a symmetry-reduced neural information field for assessing the observation geometry of Walker LEO sensor constellations, with numerical validation focused on moving targets at altitudes of 15–100 km. The reference quantity is the A-optimal potential ψA, minimized over feasible sensor subsets of a prescribed size within a range-ranked candidate pool. Smaller values indicate more informative local geometry. The reported implementation and experiments use three-satellite subsets. With standard J2 dynamics defining the nominal constellation model, the proposed reduction uses axial equivariance and a topology-dependent finite permutation to map absolute time to bounded co-precessing and fast phases. A two-stage coordinate network then approximates ψA over target longitude, latitude, altitude, and phase, providing differentiable potential estimates for feasible queries within the sampled domain. Separate tests quantify neural approximation error and field validity under matched and high-fidelity dynamics and on historical operational trajectories. The principal contributions are threefold:1.A symmetry-reduced representation maps Earth rotation, J2 nodal precession, and topology-specific Walker slot recurrence to bounded co-precessing and fast phase coordinates for the angle-only information field.2.A two-stage differentiable surrogate provides efficient pointwise estimates of the candidate-pool potential ψA and its spatial gradients for geometric assessment.3.The validity analysis distinguishes neural approximation error from dynamics–model mismatch and quantifies field validity under uncontrolled high-fidelity dynamics and along historical Iridium operational trajectories. An application demonstration evaluates field-guided search triggering for sensor reconfiguration. 1.A symmetry-reduced representation maps Earth rotation, J2 nodal precession, and topology-specific Walker slot recurrence to bounded co-precessing and fast phase coordinates for the angle-only information field. A symmetry-reduced representation maps Earth rotation, J2 nodal precession, and topology-specific Walker slot recurrence to bounded co-precessing and fast phase coordinates for the angle-only information field. 2.A two-stage differentiable surrogate provides efficient pointwise estimates of the candidate-pool potential ψA and its spatial gradients for geometric assessment. A two-stage differentiable surrogate provides efficient pointwise estimates of the candidate-pool potential ψA and its spatial gradients for geometric assessment. 3.The validity analysis distinguishes neural approximation error from dynamics–model mismatch and quantifies field validity under uncontrolled high-fidelity dynamics and along historical Iridium operational trajectories. An application demonstration evaluates field-guided search triggering for sensor reconfiguration. The validity analysis distinguishes neural approximation error from dynamics–model mismatch and quantifies field validity under uncontrolled high-fidelity dynamics and along historical Iridium operational trajectories. An application demonstration evaluates field-guided search triggering for sensor reconfiguration. 2. Theory and Symmetry Reduction The angle-only information field is determined jointly by sensing geometry and constellation dynamics. This section first formulates the angle-only information measure and a nominal Walker constellation under standard J2 dynamics. The time-dependent field is then represented on a bounded phase domain using the constellation’s axial and finite-permutation symmetries, without requiring the field itself to be smooth. 2.1. Problem Definition and Reference Frames A field query is specified by the Earth-fixed geocentric longitude λE, geocentric latitude φ, altitude h, and time t. The vector rTE denotes the geocentric position of the query target, measured from the Earth center O and resolved along the axes of the adopted Earth-fixed frame FE. Under the spherical-Earth approximation, the target position is given by(1)rTE(λE,φ,h)=(RE+h)cosφcosλEcosφsinλEsinφ, where RE is the reference Earth radius and h is the radial altitude above the reference sphere. Longitude λE is measured in the equatorial plane from the positive xE axis toward the positive yE axis. Let t0 be the reference epoch, Δt=t−t0, and θG,0 the Earth rotation angle at t0. The adopted constant-rate Earth rotation model is(2)θG(t)=θG,0+ωEΔt,rE(t)=Rz[−θG(t)]rI(t). Here, ωE is the Earth rotation rate, and the superscript I denotes the geocentric inertial frame FI. The angle θG(t) describes the positive rotation of the Earth-fixed axes relative to the inertial axes about their common z-axis. With the active-rotation convention for Rz, the inertial-to-Earth-fixed component transformation uses −θG(t). Figure 1 illustrates the reference frames and target coordinates. Consider a Walker-type LEO sensor constellation with Np orbital planes and Ns satellites per plane, giving N=NpNs sensors in total [26]. The plane and in-plane slot indices are p∈{0,…,Np−1} and k∈{0,…,Ns−1}, respectively. At the reference epoch, the node and phase relations shown in Figure 2 are(3)Ωp=Ω0+pΔΩ,ΔΩ=ΩspanNp,Mp=modM0+2πFpNpNs,2π,Mp,k=modMp+2πkNs,2π. Here, Ω0 is the reference right ascension of the ascending node (RAAN), M0 is the reference along-track phase, Ωspan is the node span, ΔΩ is the adjacent-plane node spacing, and F is the Walker phasing parameter. The orbital size is common to all slots, and i is the common inclination. Thus, 2π/Ns is the separation between adjacent slots in one plane, whereas 2πF/(NpNs) is the phase offset between the reference slots of adjacent planes. These relations describe the conventional Walker geometry; Section 2.3 constructs the corresponding common-epoch states on a numerically closed J2 reference orbit. At time t, the propagated satellite positions are denoted by riE(t), i=1,…,N. Let P collect the fixed orbital, frame, Earth-rotation, sensing, and normalization parameters. The reference model maps a query to an information potential and a feasibility indicator,(4)Ψref:(λE,φ,h,t;P)⟼(ψA,f). The parameter set P includes the Walker geometry, the relative-periodic reference orbit, the frame convention, and the sensing constraints. The potential ψA and feasibility indicator f are defined below. 2.2. Angle-Only Information Potential The reference calculation first determines which satellites can provide admissible angle measurements. For satellite i, define the target-to-sensor relative position, range, and unit line-of-sight vector by(5)ΔriE≡riE−rTE,ρi≡∥ΔriE∥2,ℓi≡ΔriEρi. Earth occlusion is evaluated on the complete target-to-sensor segment. Its minimum geocentric distance is(6)diseg=min0≤u≤1rTE+uΔriE2. For the reference field used in this study, the visible set is(7)V=i:ρi≤ρmax,diseg>RE+cE, where ρmax is the maximum sensing range and cE is the required Earth clearance. The measurement model assumes locally isotropic, independent angular errors with standard deviation σθ in the two-dimensional tangent plane of each line of sight. Linearization with respect to the target position gives the single-sensor Fisher information matrix (FIM)(8)Ji=1σθ2ρi2I3−ℓiℓiT. The projection matrix in Equation (8) contributes information perpendicular to the line of sight but none along it. The range-dependent weight wi=(σθ2ρi2)−1 therefore favors nearby measurements, and the sum of several FIMs captures whether their viewing directions complement one another. This use of Fisher information is consistent with information-based angle-only sensing and sensor-selection formulations [5,6]. The reference criterion is defined on a range-weighted candidate set. Let Nsel≥2 denote the prescribed number of selected sensors; the implementation and all experiments reported here use Nsel=3. The visible sensors are ranked by wi, and(9)K=TopK0V;wi contains the K0 largest-weight sensors, or all visible sensors when |V|