acceptodds
Under review as a conference paper at ICLR 2027

MrSOP: A Multi-Regime Satellite Orbit Prediction Model Based on Tensor Decomposition

Abstract

Accurate satellite orbit prediction is critical for collision avoidance and sustainable space operations. Existing learning-based methods, however, are typically trained on few satellites or a single orbital regime. We construct a Two-Line Element (TLE) dataset of 18,004 satellites across LEO, MEO, NSO, and GEO, and propose MrSOP: a single multi-regime forecaster with physics-anchored routing and a Decomposed Mixture of Experts (DMoE). It recovers semi-major axis and eccentricity from the input, and needs no separate learned gating network or orbital-regime label at inference. MrSOP attains the lowest error in all eight in-domain regime–metric comparisons, reducing MAE by up to 60.2% over the second-best model. It achieves the lowest error among learned models in all six zero-shot source–metric comparisons. A 16-expert DMoE cuts parameters by 94.2% relative to the full-rank model, with a gain in mean in-domain accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
Satellite Orbit Prediction, Time-series Forecasting, Tensor Decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3GijOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MrSOP, a single multi‑regime orbit forecasting model combining physics‑anchored routing and tensor‑train decomposed mixture‑of‑experts. A large‑scale TLE‑derived dataset covering LEO/MEO/NSO/GEO is constructed. It achieves strong zero‑shot forecasting results on real held‑out satellite precise ephemerides with greatly reduced parameter count via DMoE. The work demonstrates promising application‑level progress for satellite orbit prediction, though several gaps remain between the experimental setup and real‑world space‑situational‑awareness deployment.

Targeting satellite orbit prediction across multiple orbital regimes including LEO, MEO, NSO and GEO is a relatively novel angle. Based on this merit, I am inclined towards a positive evaluation. Nevertheless, the authors must provide reasonable responses to the concerns I have raised.

Critical strengths
  1. The paper targets multi‑regime satellite orbit prediction covering LEO, MEO, NSO and GEO within a single model, which is a relatively novel setting. Its Mixture‑of‑Experts‑based (named DMoE, i.e. Decomposed Mixture‑of‑Experts) design is explicitly motivated by heterogeneous orbital dynamics across different regimes: the MoE backbone provides a pool of experts to fit distinct dynamical patterns, and physics‑anchored routing dispatches inputs to suitable experts using physically derived orbital‑element coordinates. The “Decomposed” part refers to tensor‑train weight decomposition purely for parameter compression and does not alter the multi‑regime routing logic. This avoids training separate models for each orbital regime and does not require regime labels at inference time.

  2. A large‑scale TLE‑derived dataset of 18,004 satellites across four orbital regimes is carefully constructed. Critically, zero‑shot evaluation is performed on held‑out real‑world precise ephemerides completely unseen during training, which improves experimental rigor compared with many orbit‑prediction works.

Critical weaknesses
  1. The dataset suffers from severe imbalance in satellite counts among different orbital regimes. Despite DMoE seeming to be a promising approach for multi‑regime satellite orbit prediction, several issues remain unaddressed. These include the criterion for selecting the number of experts, the mitigation strategy for cross‑regime data‑distribution imbalance, and the justification for hard‑coding activation of two experts per input window rather than 1, 3 or alternative values.

  2. DMoE achieves massive parameter compression accompanied by noticeable accuracy improvement instead of performance drop. This can be attributed to either overfitting in the original full‑rank MoE or implicit low‑rank regularization brought by tensor‑train decomposition. Unfortunately, the paper only reports empirical metrics, without diagnostic experiments to separate these two effects or supporting theoretical analysis.

Questions for authors
  1. Different orbital regimes are governed by distinct dominant perturbation forces. How does your model handle fundamentally different dynamics, given that you only route via (log\ a,e) and do not embed explicit perturbation physics inside experts? Why does it fail on some GEO zero‑shot samples?
  2. The paper fixes temperature \tau=0.15 and top‑2 expert selection. There is no ablation study varying \tau or the number of activated experts per forward pass. These are critical heuristic design choices; how sensitive is performance to these hyperparameters?
  3. Why top‑2? Would top‑1 or top‑3 or others yield better or worse forecasting quality, especially for boundary orbital states between regimes?
  4. The dataset is heavily imbalanced: LEO dominates sample count, while MEO, NSO and GEO contain far fewer trajectories. Could the large volume of LEO data dominate training? Can MEO/GEO‑related experts learn meaningful patterns with limited gradient signals?
  5. DMoE improves accuracy after heavy parameter compression. Is this caused by full‑rank MoE overfitting or implicit low‑rank regularization from tensor‑train decomposition? Could you provide experiments or analysis to tell them apart?
3 comments

We thank the reviewer for recognizing the unified multi-regime design and the evaluation on unseen ephemerides.


W1. Expert configuration and regime imbalance

Expert count. In Appendix E (Effect of the Number of Experts, lines 1109–1120), sixteen experts give the lowest Starlink and GNSS errors among the tested expert counts while using fewer parameters than 32 experts. Varying E over 4, 8, 16, and 32 gives in-domain Pos3D MAE of 151.35–160.40 km (E denotes K in the paper). Increasing from 16 to 32 raises parameters from 1.96 M to 3.59 M without improving either zero-shot result, supporting 16 as the common configuration.


W2 & Q5. Why TT improves accuracy under compression

Why accuracy improves under compression. Table A supports a contribution from TT's implicit low-rank regularization. TT constrains the expert weights while retaining full-width experts. Compared with the original full-rank MoE, TT has comparable training MAE and lower validation and test errors.

Matched parameter budgets. TT achieves 192.2 km test Pos3D MAE with 1.96 M parameters and expert width 1,024. The narrow full-rank control achieves 197.6 km with 1.97 M parameters and width 52. TT also has lower training and validation errors than this control.

Compression and rank choice. The original full-rank MoE requires 33.85 M parameters and has 248.7 km test error. Among TT ranks 8, 16, 32, and 64, rank 32 gives the lowest validation and test errors.

Table A. Rank sweep and parameter-matched control. Train and Val MAE average errors over six channels in normalized units. Test error is Pos3D MAE (km). In all comparison tables, bold marks the lowest error and underlining marks the second-lowest.

ModelParamsTrain MAEVal MAETest Pos3D MAE (km)
Full-rank MoE (E = 16)33.850 M2.700e-2‾\underline{\text{2.700e-2}}1.930e-2248.700
Full-rank MoE, parameter-matched (dffd_{\mathrm{ff}} = 52)1.970 M2.890e-21.870e-2197.600‾\underline{\text{197.600}}
TT rank 80.510 M3.080e-22.250e-2305.400
TT rank 161.130 M2.670e-21.930e-2213.400
TT rank 32 (default)1.960 M2.670e-21.820e-2192.200
TT rank 643.640 M2.670e-21.840e-2‾\underline{\text{1.840e-2}}203.500

We will add the rank sweep and parameter-matched comparison to Section 5.2, with training and validation diagnostics in the appendix.


Q1. Different orbital dynamics and GEO zero-shot performance

Handling different orbital dynamics. MrSOP uses (log⁡a,e)(\log a,e) to select and weight experts, while each expert predicts from the position–velocity history. Experts learn temporal patterns from these histories and can capture the signatures of perturbations represented in the training trajectories. Inputs with similar routing coordinates can therefore produce different forecasts. Shared TT cores support learning across regimes.

Why route with (log⁡a,e)(\log a,e)? We use (log⁡a,e)(\log a,e) as a compact physical coordinate pair for expert routing. Semi-major axis and eccentricity are established descriptors of orbital size and shape [1]. They determine the perigee and apogee heights used in ESA's altitude-based definitions of LEO, MEO, and HEO [2, Table 1.2]. Their use follows established aa–ee dynamical maps at specified inclinations [3, Section 3.3]. We use log⁡10(a/km)\log_{10}(a/\mathrm{km}) to reduce scale differences between low and high orbits before standardization. These coordinates guide expert selection, while each expert predicts from the full position–velocity history.

Additional routing coordinates. Other orbital elements can refine distinctions within a regime. Inclination, for example, enters ESA's GEO and NSO definitions [2] and affects long-term lunisolar dynamics [4]. Section 6 (Discussion, lines 518–520) identifies inclination as a future routing coordinate. The current experts retain orbit-specific information through the full state history.

GEO zero-shot performance. Similar semi-major axes and near-zero eccentricities can make GEO trajectories difficult to distinguish in (log⁡a,e)(\log a,e) space. This may limit routing specialization within GEO and contribute to lower accuracy on unseen GEO satellites. Inclination is not an explicit routing input, so small orbital-plane differences among near-equatorial trajectories are represented through the full position–velocity histories.

Performance across regimes. Nevertheless, in the paper's main comparison (Table 3), MrSOP achieves the lowest BeiDou error among learned models and ranks second to SGP4 on this eight-satellite GEO cohort (Appendix B.3). Across the main Tables 2–3, MrSOP achieves the lowest errors on Starlink, GNSS, and all four in-domain regimes, using one model without regime-specific retraining or regime labels at inference.

We will add this coordinate-selection rationale and its supporting references to Section 3.2 (Physics-anchored Routing).

Q2 & Q3. Routing sensitivity and the choice of two active experts

Default configuration and sensitivity. Our default configuration, τ=0.15\tau=0.15 with top-2 routing, gives the lowest test-average error and the lowest seven-set average in both separate sensitivity sweeps. The seven-set average is 175.957 km, supporting one common configuration across regimes. Changing τ\tau to 0.05 or 0.5 increases the test-average error by at most 3.3%. Using top-1, top-3, or top-4 increases it by at most 4.2%.

Routing temperature. τ=0.15\tau=0.15 gives 192.2 km test Pos3D MAE, compared with 198.6 km at 0.05 and 198.0 km at 0.5 (Table B).

Table B. Routing temperature: Pos3D MAE (km), 10-epoch budget, seed 42. Avg. is the arithmetic mean of the seven reported regime/dataset errors, with equal weights and excluding Test: all.

τ\tauAvg.Test: allLEOMEONSOGEOZero-shot: StarlinkZero-shot: GNSSZero-shot: BeiDou
0.050187.214198.600199.600197.600152.700‾\underline{\text{152.700}}243.90097.100‾\underline{\text{97.100}}241.200‾\underline{\text{241.200}}178.400
0.150 (default)175.957192.200200.000‾\underline{\text{200.000}}206.400‾\underline{\text{206.400}}139.200222.700‾\underline{\text{222.700}}96.700204.600162.100‾\underline{\text{162.100}}
0.500178.171‾\underline{\text{178.171}}198.000‾\underline{\text{198.000}}204.100227.100186.700174.60098.200242.600113.900

Activated experts. Top-2 gives 192.200 km test Pos3D MAE, compared with 199.400, 200.200, and 197.450 km for top-1, top-3, and top-4, respectively (Table C).

Table C. Activated experts per window (top-kk): Pos3D MAE (km), 10-epoch budget, seed 42. Avg. uses the same seven-set arithmetic mean as Table B.

Top-kAvg.Test: allLEOMEONSOGEOZero-shot: StarlinkZero-shot: GNSSZero-shot: BeiDou
1178.486‾\underline{\text{178.486}}199.400206.100230.600143.100217.80092.100208.800150.900
2 (default)175.957192.200200.000206.400139.200222.700‾\underline{\text{222.700}}96.700‾\underline{\text{96.700}}204.600‾\underline{\text{204.600}}162.100‾\underline{\text{162.100}}
3182.614200.200208.900244.900106.800239.800110.600198.100169.200
4183.680197.450‾\underline{\text{197.450}}202.690‾\underline{\text{202.690}}222.550‾\underline{\text{222.550}}128.200‾\underline{\text{128.200}}235.890100.240215.620180.570

Why top-2, including boundary states? Top-2 combines the lowest test-average error in this sweep with the ability to blend two expert representations near routing boundaries. Individual regimes favor different settings. Table D under Q4 reports aggregate expert utilization. The ablations do not separately measure forecast errors on boundary windows.

We will add the temperature and top-kk ablations to Appendix D and summarize their findings in Section 5.2 (Ablation Study).


Q4. Regime imbalance and learning in smaller regimes

Learning across unequal regimes. Square-root sampling assigns LEO a sampling probability of 49.7%, compared with its 70.4% share of available windows. Under the same total training-window budget, MrSOP also outperforms separately trained models in the smaller regimes, including MEO and GEO.

The sampler draws regimes in proportion to the square root of their window counts, raising MEO's sampling probability from its 1.4% data share to 6.9%.

Learning in smaller regimes. Routing weights experts by distances to anchors in (log⁡a,e)(\log a, e) space, while shared TT cores receive training signals across regimes. Table D shows distinct dominant experts for LEO, NSO, and GEO, with MEO using several experts.

Table D. Expert utilization by regime for MrSOP with seed 42. Values are shares of the regime's routed gate weight in the last MoE layer.

RegimeExperts receiving > 1% of the regime's gate weight# experts above 1%
LEOe3 (100.000%)1
MEOe6 (85.300%), e2 (5.700%), e11 (4.600%), e14 (3.700%)4
NSOe11 (100.000%)1
GEOe13 (100.000%)1

Per-regime evidence. Under the same total training-window budget, MrSOP outperforms four separately trained TimeMixer models in every regime, including MEO (206.4 vs. 360.3 km) and GEO (222.7 vs. 475.1 km), measured as test Pos3D MAE with a 10-epoch budget and seed 42. TimesFM [5] and Chronos-2 [6] also learn from heterogeneous collections of unequal size. Our sampling and per-regime results provide the evidence for MrSOP.

We will specify the sampling probabilities in Section 4 and add the expert-utilization and per-regime comparisons to Appendix D.

References

[1] Montenbruck, O., and Gill, E. Satellite Orbits: Models, Methods and Applications. Springer, 2000. https://doi.org/10.1007/978-3-642-58351-3.

[2] ESA Space Debris Office. ESA's Annual Space Environment Report, Issue 10.1, 2026. https://www.sdo.esoc.esa.int/environment_report/Space_Environment_Report_latest.pdf.

[3] Rosengren, A. J., Skoulidou, D. K., Tsiganis, K., and Voyatzis, G. Dynamical cartography of Earth satellite orbits. Advances in Space Research, 63(1), 443–460, 2019. https://doi.org/10.1016/j.asr.2018.09.004.

[4] Daquin, J., Rosengren, A. J., Alessi, E. M., Deleflie, F., Valsecchi, G. B., and Rossi, A. The dynamical structure of the MEO region: long-term stability, chaos, and transport. Celestial Mechanics and Dynamical Astronomy, 124(4), 335–366, 2016. https://doi.org/10.1007/s10569-015-9665-9.

[5] Das, A., Kong, W., Sen, R., and Zhou, Y. A decoder-only foundation model for time-series forecasting. Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 10148–10167, 2024. https://proceedings.mlr.press/v235/das24c.html.

[6] Ansari, A. F., et al. Chronos-2: From Univariate to Universal Forecasting. arXiv:2510.15821, 2025. https://arxiv.org/abs/2510.15821.

Reviewer_nfR9OpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a forecasting model that is designed to work for satellites in any orbit, across LEO, MEO and GEO. Starting from TLE data, it constructs a dataset in cartesian space to train the model. The model itself is TimeMixer with MoE. It shows improvement against other learning-based methods in in-distribution, and also for and out-of-distribution case, using ephemeris data, which is far more accurate than TLE data.

Critical strengths

The question of multi-regime orbit prediction is relevant, and the authors gather a significant number of data. The authors recognize some of the limitations in the results, namely the SGP4 data and that tsFM are able to surpass MrSOP if finetuned, for some data.

Critical weaknesses
  1. The authors failed to compare with previous work in orbit prediction ( as an example, AutoOrbit: Physics-Informed Satellite Orbit Prediction, deltaSGP4,… ). Also please cite the RevIN paper.

  2. Figure 2 a) is blatantly wrong. That is not a feasible orbit, and if the authors use this as ground truth it explains the strange results.

  3. The algorithmic improvement is limited, and the use of altitude + eccentricity to separate the anchors is not relevant for GEO satellites, which have eccentricty = 0, inclination = 0 and all have the same altitude. The authors correctly recognize this in the limitations.

Questions for authors

(1) please explain in more detail the interpolation used in the paper, including the propagation model used to fill between tle.

(2) The results for SGP4, which is a common method in the industry are suprising, specially for the zero-shot ephemeris dataset. Claiming that SGP4 has an MAE of 20.000km when predicting Starlink data is a big claim. The authors should analyze and careful explain why this happen. Also what does this mean? “because every sampled file is paired with a TLE about 77 days older than the ephemeris.”

(3) please add a numerical propagator to the comparison.

(4) can you add a plot or table showing the error over time.

(5) can you provide a specific evaluation for periods of increased air density, during solar flares?

5 comments

We thank the reviewer for recognizing the relevance of multi-regime orbit prediction and the dataset scale.


W1 & Q3. Orbit-prediction baselines and practical reliability

Practical reliability across regimes. The propagators' lower BeiDou errors alone do not establish dependable forecasting across all regimes. Under our archived-TLE protocol, forecasts were unavailable for 1,279 requested windows with dSGP4 and 1,535 with Orekit (Table C). MrSOP requires neither archived TLEs nor external ephemerides at inference, while giving lower reported errors across all four in-domain regimes and on Starlink/GNSS (Tables A–B).

AutoOrbit reproducibility [1]. The released code does not support a faithful AutoOrbit comparison in our setting. The official release omits the original HPOP implementation and precomputed celestial data, and its physics-loss interface returns zero with a skipped status. Our FNO-based adaptation changes the reference-orbit construction and omits this objective, so its scores are excluded from Tables A–B.

Evaluation setting. AutoOrbit predicts residuals relative to a reference orbit and provides optional maneuver correction using maneuver records. Our benchmark evaluates 24-hour position–velocity forecasts across LEO, MEO, NSO, and GEO from 48-hour state histories, without maneuver inputs. A comparable implementation requires restoring the missing physics components and validating the adaptation to this protocol, which we could not complete within the rebuttal period.

dSGP4 [2]. We used the official differentiable SGP4 implementation with the same archived TLEs as SGP4.

RevIN citation. We have added the RevIN citation [3] in Section 3.1 (lines 200–201), where normalization is introduced.

Orekit [4]. The numerical propagator starts from the same TLE's SGP4 state at the forecast origin. It integrates the EIGEN-6S 8×8 geopotential and Sun/Moon third-body gravity with Dormand–Prince 8(5,3), using the existing 1 m position tolerance.

Matched evaluation protocol. We evaluated dSGP4 and Orekit across all 116,287 designated windows using the original SGP4 protocol. We reused the original in-domain windows and all Starlink, GNSS, and BeiDou windows.

Control and aggregation. dSGP4/Orekit scores in Tables A–B use windows where all three propagators succeeded. Table C counts unavailable forecasts against all requested windows.

Table A. In-domain Pos3D MAE (km). dSGP4/Orekit means use common-success windows. In all comparison tables, bold marks the lowest error and underlining marks the second-lowest.

RegimedSGP4OrekitMrSOP
LEO8,900.1498,899.835‾\underline{\text{8,899.835}}175.700
MEO1,028.625‾\underline{\text{1,028.625}}1,029.439178.700
NSO130.784‾\underline{\text{130.784}}130.834117.900
GEO2,487.252‾\underline{\text{2,487.252}}2,487.581149.000
All3,152.034‾\underline{\text{3,152.034}}3,152.252155.300

Table B. Zero-shot evaluation on the three ephemeris datasets used in the paper (Pos3D MAE, km). Propagator means use common-success windows.

DatasetdSGP4OrekitMrSOP
Starlink77,858.93274,067.729‾\underline{\text{74,067.729}}84.600
GNSS419.742‾\underline{\text{419.742}}419.837180.400
BeiDou2.1764.446‾\underline{\text{4.446}}110.900

Table C. Unavailable forecasts (NaN), including input and propagation failures.

DatasetdSGP4 NaNOrekit NaNInput gaps (both)
In-domain16843732
Starlink974961675
GNSS137137137
BeiDou000
All1,2791,535844

Unavailable forecasts. Across 116,287 requested windows, dSGP4 and Orekit had 1,279 (1.10%) and 1,535 (1.32%) unavailable forecasts. Both totals include 844 missing-input windows: 812 without TLEs and 32 without timestamps. The SGP4 control also had 1,279 unavailable windows.

Failure categories. Beyond these input gaps, dSGP4 reported 435 propagation errors or nonfinite states (136 in-domain LEO and 299 Starlink). Orekit could not obtain an initial SGP4 state for 421 windows (135 in-domain LEO and 286 Starlink). Another 270 MEO windows fell outside the configured DE440 dates and raised “no data generated around date.” This last category is an auxiliary-data coverage failure.

Practical implication. Both evaluated propagator configurations depend on valid TLE initialization, and Orekit additionally depends on date-compatible ephemerides. They cannot provide complete forecast coverage under this protocol without resolving the observed failures. MrSOP uses one model across regimes and forecasts directly from recent position–velocity histories, avoiding these auxiliary-input dependencies.

We will report both forecasting accuracy and unavailable forecasts in Section 5.1 (Main results), with propagation settings and failure categories in Appendix F.

W2. The orbit shown in Figure 2(a)

Ground truth versus visualization. The polygonal appearance in Figure 2(a) comes from straight lines connecting positions sampled at 15-minute intervals. The Starlink reference states come from held-out operational ephemerides. The connecting segments are not ground-truth states. Forecasting errors are evaluated at the sampled epochs.

Appendix illustration. We clarified the comparison of MrSOP predictions with ground truth at matching forecast times in Section 5.1 (lines 428–431) and the Figure 2 caption (lines 447–451). The new illustration in Appendix B.4 (lines 987–999, Figure 5) overlays the 15-minute positions on the 1-minute operational ephemeris, showing that the sampled positions lie on the reference orbit.


W3. Algorithmic contribution and routing for GEO

GEO routing and prediction. We agree that GEO samples occupy a narrow region in (log⁡a,e)(\log a,e), which can limit the router's ability to distinguish variation within this regime. MrSOP nevertheless produces satellite-specific forecasts because each expert processes the individual position–velocity history. Similar routing weights therefore do not imply identical predictions.

Actual GEO dynamics. Zero eccentricity and inclination describe an ideal geostationary orbit. In operation, perturbations change these elements, while station-keeping controls their departures from target values [12]. In a deterministic audit of 100 GEO input windows, both quantities were nonzero at the last observed state of every window. MrSOP derives (log⁡a,e)(\log a,e) from the observed states, preserving their orbital variation. The three-dimensional state histories also retain orbital-plane geometry, while adding inclination to the router remains a future direction in Section 6 (lines 518–520).

Algorithmic contribution. Our contribution combines orbit-dependent routing with shared TT experts across regimes. With the same 10-epoch budget and seed 42, top-2 routing gives 192.200 km test Pos3D MAE, compared with 199.400 km for top-1 and 200.200 km for top-3.

We will clarify the distinction between routing coordinates and trajectory prediction in Sections 3.2 and 3.4.


Q1. Propagation used to reconstruct trajectories between TLEs

Propagation used to fill between TLEs. The intermediate states are generated by orbital propagation, so they retain the physical effects modeled by the propagator. For the in-domain dataset, we use an Orekit-based pipeline for LEO and SDP4 for MEO, NSO, and GEO. Orekit's TLEPropagator implements SGP4/SDP4 propagation from TLE epochs and orbital elements [4]. These models account for Earth's gravity and atmospheric drag. SDP4 additionally includes solar/lunar perturbations and orbital resonances [5]. We generate one-minute states, apply the Section 2 filtering, and retain every 15th state. TLE gaps over 72 hours split trajectory parts, which forecasting windows never cross.

What MrSOP learns. The reconstructed states already reflect the physical effects included in the propagator. MrSOP is therefore supervised on trajectories containing those effects and can learn their temporal signatures without receiving each underlying physical variable separately. Appendix B.2 compares the reconstructed trajectories with independent ephemerides.

We will expand the propagation-model description in Section 2.1 and Appendix B.1.

Q2. Large SGP4 errors and the 77-day TLE age

Why the SGP4 errors are large. The Starlink mean reflects substantial phase offsets and a small number of extreme finite predictions under the archived-TLE protocol. Figures 6–7 in Appendix F (OneWeb and Starlink, lines 1180–1184) show SGP4 trajectories with similar periods and coordinate amplitudes to the reference but a phase shift already at the first forecast step. Comparing positions at the same timestamps therefore yields large errors despite the similar curve shapes. Across 9,462 windows with finite SGP4 metrics, the median Pos3D MAE is 6,686.454 km. However, 37 windows (0.391%) exceed 100,000 km and contribute 91.166% of the summed position error, raising the mean to 77,858.932 km. Replaying the worst Starlink case also confirms that SGP4 produces a plausible orbital radius at the TLE epoch but an unphysical radius after approximately 77 days of propagation, without an error flag.

Meaning of the reported 20,000. The reported 20,009.841 is the original six-channel MAE, averaging position (km) and velocity (km/s) errors. It is not a 3D position distance. The Pos3D mean above provides the corresponding physical distance, and the extreme finite predictions explain why that mean is much larger than the median.

Meaning of the 77-day gap. TLE age is the interval from the selected TLE epoch to the last input timestamp, distinct from the 24-hour forecast horizon. We select the latest available preceding TLE for the same satellite. Most selected Starlink TLEs date from April 8, 2026, while the last input states are from June 25. The gap follows the archive's temporal coverage rather than an imposed delay. Among the 9,462 windows, the minimum age is 77.065 days and the median is 77.288 days.

MrSOP forecasts directly from 192 recent position–velocity states, avoiding dependence on a separately matched archived TLE. We have clarified the archived-TLE ages and the effects of phase offsets and extreme errors in Appendix F (lines 1144–1150 and 1219–1221). The metric units are clarified in the captions of Tables 2–3 and 12 (lines 379–380, 395–396, and 1205–1206).


Q4. Error over forecast lead time

Error over the forecast horizon. LEO position error increases from 60.9 km at one hour to 260.4 km at 24 hours, while MEO and GEO vary non-monotonically at intermediate lead times. Table D provides the requested breakdown by regime.

Table D. Pos3D MAE (km) by lead time

Set1 h3 h6 h12 h24 h
Test: LEO60.90096.600133.900184.300260.400
Test: MEO152.800177.400176.300166.600235.500
Test: NSO99.000111.000121.800123.700189.000
Test: GEO108.700120.600141.900136.700253.000
Zero-shot: Starlink30.20041.90049.80078.800165.100
Zero-shot: GNSS134.600159.800183.900185.900302.000

Table D reports errors at individual lead times, averaged over three seeds. The main aggregate metrics average errors over all 96 forecast steps.

We will add the lead-time results in Table D to Section 5.1 (Main results).


Q5. Increased atmospheric density and space-weather inputs

Future evaluation during increased air density. We will evaluate MrSOP specifically during elevated-density periods as future work included in the revised Discussion (Section 6, lines 525–533). Orbital histories will be aligned with F10.7 and Kp indices, solar-flare records, and thermospheric density estimates to compare low and high space-weather activity and representative geomagnetic storms [6–8]. We will also test whether adding these inputs improves accuracy.

Filtering and robustness. Within comparable regimes and altitude ranges, we will check whether the 3σ3\sigma filter disproportionately rejects elevated-density intervals. Retained and rejected trajectories will be validated against independent ephemerides before evaluating robustness on valid trajectories from these intervals.

Why we use orbital histories. MrSOP learns forecasting relationships from position–velocity histories, providing a common input across regimes without requiring synchronized environmental data. Prior work similarly learns from tracked states [9], historical TLEs [10], or orbital-element histories without drag or radiation coefficients [11]. We thank the reviewer for suggesting the complementary space-weather evaluation.

Reviewer_y1DuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MrSOP, a unified satellite orbit prediction model for LEO, MEO, NSO, and GEO. Built on TimeMixer, it uses semi-major axis and eccentricity for physics-anchored expert routing and Tensor-Train decomposition to share and compress expert weights. The authors construct a TLE-derived dataset covering approximately 18,000 satellites and evaluate 24-hour predictions from 48-hour histories, including evaluation on external ephemeris datasets.

Critical strengths
  1. Comparisons cover comprehensive baselines.

  2. A large scale satellite orbit is collected.

Critical weaknesses
  1. Writing and organization could be improved. Some motivation and results are not well aligned.

  2. Application scenarios regarding the current task setting are not clear

Questions for authors
  1. What are the benefits of tensor decomposition beyond parameter reduction? If inference efficiency is a main concern, the computational cost analysis should be included in the main text. What inference speed is practical for different applications and orbital regimes?

  2. The 6-channel MAE directly averages position errors (km) and velocity errors (km/s), yet reports the result in km. This metric lacks consistent physical units. The evaluation also lacks a direct assessment of the physical consistency of predicted trajectories.

  3. Why use MoE rather than training separate models for different orbital regimes? Does MoE offer advantages in prediction accuracy, training cost, or deployment efficiency?

  4. The same prediction horizon may represent very different settings across orbital regimes and downstream tasks. What application scenarios motivate this choice? Different orbital periods and longer-term periodic effects (e.g., lunar perturbations) do not appear to be explicitly considered. How does the model account for these differences?

3 comments

We thank the reviewer for recognizing the baseline coverage and dataset scale.


W1. Presentation and alignment of claims with results

We will revise the motivation in Section 1 and the presentation of results in Section 5 so that our forecasting-accuracy and parameter-efficiency claims are tied directly to their supporting results.


W2 & Q4. Application setting, forecast horizon, and orbital dynamics

Application setting and forecast horizon. We use a 24-hour forecast for daily ground-based orbit monitoring and to cover approximately one GEO revolution. A common 15-minute cadence aligns the datasets, while temporal learning and orbit-dependent routing adapt the forecast to different orbital periods.

Sampling and context. The sampling intervals are 1 minute for the reconstructed in-domain data and Starlink, 5 minutes for BeiDou, and 15 minutes for GNSS. Their least common multiple is 15 minutes. We therefore retain the GNSS cadence and downsample the other datasets. At this interval, the 192-step input provides 48 hours of context and the 96-step output forecasts the following 24 hours, corresponding to approximately two GEO revolutions of context and one revolution of prediction.

Different orbital periods. MrSOP combines temporal pattern learning with orbit-dependent routing (Sections 3.1–3.2). Its TimeMixer backbone mixes seasonal and trend components of position–velocity histories across temporal scales. The router derives semi-major axis aa and eccentricity ee from the last input state (Eqs. (1)–(2)): aa relates to orbital period through T=2πa3/μT=2\pi\sqrt{a^3/\mu}, while ee describes orbit shape. Distances from standardized (log⁡a,e)(\log a,e) to learned anchors select and weight two experts that transform the temporal representation (Eqs. (3)–(5)). Thus, the forecast combines observed temporal patterns with expert weights conditioned on orbital period and shape.

Longer-term periodic effects (e.g., lunar perturbations). MrSOP can learn the temporal signatures of modeled perturbations from its reconstructed position–velocity histories and targets. SDP4 is the deep-space extension of SGP4 and includes modeled solar/lunar secular and periodic perturbations [1]. Orekit provides SGP4/SDP4 TLE propagation and numerical propagation. Its numerical propagators include lunar gravity through ThirdBodyAttraction [2]. Effects represented in the reconstructed trajectories therefore enter the supervision signal without separate force inputs.

Support for learning from orbital histories. These studies motivate learning orbital evolution from historical states, as MrSOP does with its Orekit- and SDP4-reconstructed trajectories. Kassas et al. [3] combine tracked-state histories with SGP4 predictions. Peng and Bai [4] learn SGP4 corrections from historical TLEs. Marie et al. [5] predict mean orbital elements from their histories without spacecraft-specific drag or radiation coefficients in their MOE-Only variant. To evaluate longer-term periodic effects, including lunar perturbations, we have added longer histories and solar/lunar ephemerides as future directions in Section 6 (lines 532–533).


Q1. Benefits of tensor decomposition and application requirements

Accuracy with fewer parameters. TT reduces the parameter count from 33.85 M to 1.96 M (17.3×) while improving mean test Pos3D MAE from 160.0 to 155.3 km (Table A). At matched parameter counts, TT also achieves 192.2 km versus 197.6 km for the narrow full-rank control (1.96 M vs. 1.97 M parameters, 10 epochs, seed 42). This supports an accuracy benefit beyond parameter reduction.

Application requirements. Our focus is parameter-efficient forecasting across regimes. MrSOP targets daily, ground-based prediction of the next 24 hours, and its current runtime fits this daily update cycle. This setting does not require onboard execution or real-time control.

Further evaluation. Lower TT ranks may improve inference speed. We will evaluate this through additional latency measurements across ranks and add the computational-cost analysis and application context to Section 5.

Q2. Physical units of the metric and physical consistency

Physically interpretable units. We will report 3D position error in km and 3D velocity error in m/s separately, and remove the km label from the mixed six-channel MAE. MrSOP also has the lowest mean errors in the four reference-relative orbital-state diagnostics reported below. Tables A–B provide the unit-separated and physical-consistency results.

Table A. Additional unit-separated evaluation: Pos3D MAE (km) and Vel3D MAE (m/s). Zero-shot rows cover the three datasets used in the paper. The 10-epoch, seed-42 controls in Q1 and Q3 are separate experiments. In all comparison tables, bold marks the lowest error and underlining marks the second-lowest. Ties at the displayed precision share the same marking.

SetTimeMixer Pos3D MAE (km)TimeMixer Vel3D MAE (m/s)Full-rank MoE Pos3D MAE (km)Full-rank MoE Vel3D MAE (m/s)MrSOP Pos3D MAE (km)MrSOP Vel3D MAE (m/s)
Test: LEO250.800267.200184.100‾\underline{\text{184.100}}197.000‾\underline{\text{197.000}}175.700188.300
Test: MEO351.700152.400159.60091.000178.700‾\underline{\text{178.700}}93.900‾\underline{\text{93.900}}
Test: NSO199.90034.100141.600‾\underline{\text{141.600}}25.600‾\underline{\text{25.600}}117.90021.100
Test: GEO326.70023.800154.700‾\underline{\text{154.700}}11.500‾\underline{\text{11.500}}149.00011.100
Test: all282.300119.500160.000‾\underline{\text{160.000}}81.400‾\underline{\text{81.400}}155.30078.700
Zero-shot: Starlink122.900137.80087.400‾\underline{\text{87.400}}98.400‾\underline{\text{98.400}}84.60095.300
Zero-shot: GNSS277.80051.700225.800‾\underline{\text{225.800}}45.400‾\underline{\text{45.400}}180.40033.900
Zero-shot: BeiDou221.40015.800133.200‾\underline{\text{133.200}}9.700‾\underline{\text{9.700}}110.9009.000

MrSOP has lower position and velocity errors than TimeMixer in all four test regimes and all three zero-shot datasets in Table A. Its mean test position error is 155.3 km, compared with 282.3 km for TimeMixer and 160.0 km for the full-rank MoE.

Trajectory consistency. MrSOP's predicted position–velocity pairs reproduce the reference orbital energy and geometry more closely than the dense controls across the four reported diagnostics (Table B). These quantities complement the separate Cartesian errors in Table A.

For position r\mathbf r (km) and velocity v\mathbf v (km/s), the standard orbital quantities are [6]:

ε=∥v∥22−μ∥r∥,h=∥r×v∥.\varepsilon=\frac{\|\mathbf v\|^2}{2}-\frac{\mu}{\|\mathbf r\|},\qquad h=\|\mathbf r\times\mathbf v\|.

a=−μ2ε,e=∥v×(r×v)μ−r∥r∥∥.a=-\frac{\mu}{2\varepsilon},\qquad e=\left\|\frac{\mathbf v\times(\mathbf r\times\mathbf v)}{\mu}-\frac{\mathbf r}{\|\mathbf r\|}\right\|.

Here, μ\mu is Earth's gravitational parameter. The quantities ε\varepsilon and hh are orbital energy and angular-momentum magnitude per unit mass. The osculating semi-major axis aa and eccentricity ee describe the instantaneous orbit's size and shape.

For a predicted quantity q^\hat q and its time-matched reference qq, relative error RelErr(q)\mathrm{RelErr}(q) is the absolute difference divided by ∣q∣|q|. Absolute error AbsErr(q)\mathrm{AbsErr}(q) measures the magnitude of the difference in the quantity's original units:

RelErr(q)=∣q^−q∣∣q∣,q∈{ε,h},AbsErr(q)=∣q^−q∣,q∈{a,e}.\mathrm{RelErr}(q)=\frac{|\hat q-q|}{|q|},\quad q\in\{\varepsilon,h\},\qquad \mathrm{AbsErr}(q)=|\hat q-q|,\quad q\in\{a,e\}.

The time-matched reference values allow for changes under perturbations [7]. Lower errors indicate closer agreement with the reference orbit's energy and geometry.

Table B. Test-set errors averaged over windows (lower is better). The four error columns correspond to RelErr(ε)\mathrm{RelErr}(\varepsilon), RelErr(h)\mathrm{RelErr}(h), AbsErr(a)\mathrm{AbsErr}(a), and AbsErr(e)\mathrm{AbsErr}(e), respectively. All errors are dimensionless except the semi-major-axis error (km).

ModelSpecific orbital energy relative errorSpecific angular momentum magnitude relative errorSemi-major axis absolute error (km)Eccentricity absolute error
MrSOP1.600e-27.980e-3273.5001.020e-2
Full-rank MoE2.040e-2‾\underline{\text{2.040e-2}}1.010e-2‾\underline{\text{1.010e-2}}353.600‾\underline{\text{353.600}}1.290e-2‾\underline{\text{1.290e-2}}
TimeMixer3.220e-21.580e-2594.7002.020e-2

MrSOP has the lowest mean error in all four reported quantities. Its mean relative errors in specific orbital energy and angular-momentum magnitude are 1.60% and 0.798%, respectively, compared with 3.22% and 1.58% for TimeMixer.

The captions of Tables 2–3 now distinguish Pos3D MAE from mixed-unit six-channel MAE (lines 379–380 and 395–396). We will add the separate position/velocity definitions and results to Sections 4 and 5.1, and the physical-consistency definitions and results to the appendix.

Q3. Why a single MoE rather than separate regime-specific models

Why MoE? A single MoE achieves lower error in every regime and reduces deployment to one 1.96 M-parameter model. The four separate models total 8.82 M parameters. Shared temporal learning and orbit-dependent routing combine common structure with specialized prediction.

Table C. Single MoE vs. four regime-specific TimeMixer models (seed 42, test Pos3D MAE in km).

SettingLEOMEONSOGEOTotal params
4 x TimeMixer (separate models)346.100‾\underline{\text{346.100}}360.300‾\underline{\text{360.300}}219.500‾\underline{\text{219.500}}475.100‾\underline{\text{475.100}}8.820 M
MrSOP (single model)200.000206.400139.200222.7001.960 M

Deployment. One checkpoint is maintained and served for all regimes, simplifying deployment compared with four regime-specific checkpoints.

We will add this comparison to Section 5.2 (Ablation Study).


References

[1] Hoots, F. R., and Roehrich, R. L. Models for Propagation of NORAD Element Sets. Spacetrack Report No. 3, 1980, Sections 2, 6, and 10. https://www.celestrak.org/NORAD/documentation/spacetrk.pdf.

[2] Orekit Development Team. API documentation: TLEPropagator (Orekit 13.0.3), https://test.orekit.org/site-orekit-13.0.3/apidocs/org/orekit/propagation/analytical/tle/TLEPropagator.html, and ThirdBodyAttraction (Orekit 7.1), https://www.orekit.org/site-orekit-7.1/apidocs/org/orekit/forces/gravity/ThirdBodyAttraction.html.

[3] Kassas, Z. M., Hayek, S., and Ahmad, J.-H. LEO satellite orbit prediction via closed-loop machine learning with application to opportunistic navigation. IEEE Aerospace and Electronic Systems Magazine, 40(1), 34–49, 2025. https://doi.org/10.1109/MAES.2024.3432105.

[4] Peng, H., and Bai, X. A medium-scale study of using machine learning fusion to improve TLE prediction precision without external information. Acta Astronautica, 204, 477–491, 2023. https://doi.org/10.1016/j.actaastro.2022.06.023.

[5] Marie, G., Ghazi, G., Revelin, B., and Dolado Perez, J.-C. Deep Learning-Based Orbit Prediction for Satellite Maneuver Detection in LEO. Proceedings of the 11th European Conference for Aeronautics and Space Sciences (EUCASS), paper 284, 2025. https://doi.org/10.13009/EUCASS2025-284.

[6] Battin, R. The Two Body Problem Continued. Lecture 2, 16.346 Astrodynamics, MIT OpenCourseWare, 2008. https://ocw.mit.edu/courses/16-346-astrodynamics-fall-2008/resources/lec_02/.

[7] NASA Jet Propulsion Laboratory. Description of Orbits and Ephemerides, section “Orbital Elements.” https://ssd.jpl.nasa.gov/orbits_doc.html.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.