acceptodds
Withdrawn from ICLR 2027

DRC: DTM-Anchored Residual Correction for Long-Horizon Thermospheric Density Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate thermospheric density forecasting is essential for a growing range of low Earth orbit (LEO) applications, including satellite formation maintenance, space-debris collision avoidance, and reentry prediction. However, learning density dynamics directly from data remains challenging, particularly under strong and rapidly varying solar and geomagnetic activity. Rather than replacing established empirical models with neural predictors, we ask a different question: can a neural network learn to correct an empirical model while retaining it as a reliable prior? We propose DRC, a DTM-anchored Residual Correction framework, which uses the Drag Temperature Model (DTM) as a persistent prior and learns a bounded residual correction from historical observations and independently encoded space-weather context. A gated mechanism adaptively controls the contribution of the learned correction, while the bounded residual limits deviations from the empirical prior. We evaluate DRC on a curated dataset covering five satellites and periods of elevated solar and geomagnetic activity. DRC consistently reduces forecasting errors compared with representative deep learning baselines, demonstrating the potential of residual learning anchored in empirical models for long-horizon scientific forecasting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific Machine Learning, Residual Learning, Physics-Informed Machine Learning, Time Series Forecasting, Space Weather

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Z4xCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a residual corrector algorithm for thermosphere density forecasting of satellite trajectories. The paper introduces an architecture which acts as a corrector on top of a DTM model (classical). The paper also implements a curated dataset including 5 satellites.

My main reason for rejecting the paper is venue fit. I think a domain journal/conference would be a much better fit. Also, I think the paper needs more work. For instance, I am not convinced the architecture does much beyond the DTM. An ablation study would quickly identify where most of the gain comes from.

Critical strengths
  • it introduces a new problem and dataset for time series forecasting.
Critical weaknesses

Major

  • I find the paper hard to follow even though I am familiar with time series forecasting.
  • Neither the problem setting nor the method are carefully explained. For instance, how does the drag temperature model
  • this problem is largely new and the ML community needs some context - what kind of time series is this, is it chaotic, why is this hard, what are the inputs, etc.
  • Given that this problem is also very specific to an application domain, I would suggest the authors submit it to a domain journal, where there would be more interest for such a paper.
  • I suspect that the heavy lifting is done by the DTM and that if the other baselines were to predict the residual as well, they might obtain a better performance. In this regard, I would like to see an ablation to conclusively prove that the architectural additions really did contribute to skill.

More:

  • what is RSE? the acronym isn't introduced.
  • also I see acronyms being used before being introduced.
  • Acronyms that aren't parsable by a large portion of the ML community shouldn't be in the title
Questions for authors
  • why the patch encoding? given that this is sequence data why not model it directly?
  • is autoregressive prediction possible in this setting? I would hardly call this time series forecasting if only a few discrete time points are predicted.
  • how much of the performance comes from your architecture and how much comes from effectively substracting the large-scale motion with the DTM? (I guess that is what it does). How do the metrics look like only on the residual? I suspect that the metrics are dominated by the DTM prediction.
Reviewer_zUXyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The method uses DTM-2020 as an atmospheric density prior, holds the latest space weather indices fixed, propagates satellite orbits, and applies a gated neural network to predict bounded log-density corrections. It is evaluated on five satellites with forecast horizons up to 24 hours. The experiments on five datasets demonstrate the effectiveness of the proposed model over six advanced time series models.

Critical strengths
  1. The paper constructs a unified dataset covering five LEO satellites.
  2. The experiments show the effectiveness of the correction model.
  3. The paper is well written and easy to understand.
Critical weaknesses
  1. The evaluation lacks a raw DTM-2020 baseline, making the contribution of neural correction unclear. Comparisons with simple bias correction, Kalman filtering, and DTM-informed neural baselines are needed to distinguish learned improvements from the benefits of the empirical prior.
  2. Another concern is that is the widely-used correction operation, the authors should answer why the correction is needed while not using the directly modeling it. And the authors should analyze how the correction model improves performance across each sample.
  3. The correction boundary bb is not provided, and the effects of the gating mechanism and the boundary have not been independently verified. The authors should provide the corresponding experiments.
  4. Results from five sets of satellites are not equivalent to cross-satellite generalization. The authors should conduct separate cross-domain experiments.
  5. All the baselines are deep learning time series models; the authors should include the domain-specific baselines for analysis and comparison.

Additional minor issues: The MSE/RSE for GRACE-FO are inconsistent with the given formula.

Questions for authors
  1. The experiment results of Informer on the GOCE dataset are very strange; please check the experimental settings and ensure the baseline has been well trained.
  2. Please reply to the weaknesses listed above.
Reviewer_QuNrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper forecasts thermospheric mass density along satellite orbits up to 24 h ahead. Instead of predicting the density directly, a Transformer conditioned on space-weather indices learns a bounded correction to the empirical model DTM-2020. The authors also build a multi-satellite dataset (CHAMP, GOCE, GRACE, GRACE-FO) and show that the model outperforms generic deep time-series forecasters, in particular at long horizons.

Critical strengths
  • Anchoring the forecast on an empirical model, using only information that is available at the forecast origin, and bounding the correction is a reasonable way to avoid the drift of purely data-driven forecasters. The error of DRC grows little with the horizon (Fig. 5/6).
Critical weaknesses
  • Probable test leakage through the prior. DTM-2020 (Bruinsma & Boniface, 2021) was fitted to CHAMP (2001-2010), GRACE (2002-2016) and GOCE (2009-2013) densities. With chronological 80/10/10 splits, the test segments of these satellites lie inside the DTM fitting data. The results fit this explanation: the large gains are on CHAMP and GRACE-A, while on GRACE-FO, the only satellite whose test period is after the DTM data, DRC does not beat PatchTST/DLinear in horizon-averaged MAE/MSE (Table 2). This may explain most of the reported improvements.

  • Missing baselines and limited ML contribution. All baselines are generic forecasters that do not get the prior, so the comparison is essentially "with prior" vs. "without prior". Missing are:

    • DTM-2020 alone (the zero-correction model),
    • a classical calibration, e.g. a scale factor from the last orbit(s) (Doornbos et al., 2008),
    • other empirical models (NRLMSIS 2.x, JB2008) and ML density models from the space-weather literature (e.g. MSIS-UQ, Acciarini et al., 2024),
    • the generic forecasters given the DTM prior as input or trained on the residual.

    The method itself is a PatchTST with an additive residual on top of an empirical model, which is not new (e.g. APHYNITY, MSIS-UQ). So the contribution rests on the evaluation, which is the weak part. I think the paper fits better into a domain journal like Space Weather or JSWSC.

Questions for authors
  1. What are the errors of DTM-2020 alone (same persistence indices, propagated orbit) and of DTM-2020 with a last-orbit scale-factor calibration on each test set?
  2. What are the exact test periods per satellite, and how do they overlap with the DTM-2020 fitting data? Can you report results only on non-overlapping periods?
  3. Are the "latest available" indices definitive values (only available with delay) or real-time values? Do daily indices (ApAp, F10.7F_{10.7}) of the current day leak future information at 5-min resolution?
  4. How is the orbit propagated? If propagated from precise orbits with a perfect force model, the positions are near-truth.
  5. How were the bound bb and the baselines tuned? All results are from a single seed; please report mean/std over several seeds.

Minor points / Details

  • Table 3: the "Transformer-based backbone" equals PatchTST in Table 1 (0.2631/0.1457/0.5595). Is the backbone PatchTST?
  • Some baselines have RSE > 1 (e.g. DLinear on GRACE-A, 1.16), i.e. worse than predicting the test mean. Were they trained correctly and is normalization consistently applied?
  • Fig. 1(b): the radar chart shows only five of the six baselines.
  • Eq. 7: the MSE in linear standardized density is dominated by solar-maximum / low-altitude periods. A log-space loss or density ratios (standard in the field) would be more appropriate.
  • App. A: keeping only the first observation in each 5-min bin discards most of the data; why not average?
  • "Long-horizon" for 24 h is a stretch; operational forecasts need up to 3 days. Evaluation during storms and with forecasted drivers would be more informative.
  • No statement about code/data release, although the dataset is claimed as a contribution. Also, what does it add over the STORM-AI benchmark?
  • L.124: the citations for ETT/Weather/Electricity benchmarks look misplaced.
  • Related work could be more compreshensive on density-model calibration and ML, e.g.
    • E. Doornbos, H. Klinkrad, P. Visser. Use of two-line element data for thermosphere neutral density model calibration. Advances in Space Research 41(7):1115-1122, 2008.
    • E. Forootan et al. Forecasting global and multi-level thermospheric neutral density and ionospheric electron content by tuning models against satellite-based accelerometer measurements. Scientific Reports 12, 2095, 2022.
    • G. Acciarini, E. Brown, T. Berger, M. Guhathakurta, J. Parr, C. Bridges, A. G. Baydin. Improving Thermospheric Density Predictions in Low-Earth Orbit With Machine Learning. Space Weather 22, e2023SW003652, 2024.
    • S. A. Malik, J. Walsh, G. Acciarini, T. E. Berger, A. G. Baydin. High-Cadence Thermospheric Density Estimation enabled by Machine Learning on Solar Imagery. arXiv:2312.06845, 2023.
    • R. J. Licata, P. M. Mehta, D. R. Weimer, W. K. Tobiska, J. Yoshii. MSIS-UQ: Calibrated and Enhanced NRLMSIS 2.0 Model With Uncertainty Quantification. Space Weather 20(11), e2022SW003267, 2022.
    • D. J. Gondelach, R. Linares. Real-Time Thermospheric Density Estimation via Two-Line Element Data Assimilation. Space Weather 18, e2019SW002356, 2020.

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

Related papers

Discussion (0)

Sign in to comment.