Pith. sign in

REVIEW 4 major objections 6 minor 25 references

Learning What Matters: Causal Time Series Modeling for Arctic Sea Ice Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that selecting predictors by causal discovery (MVGC or PCMCI+) before training a hybrid GRU-LSTM yields Arctic sea-ice-extent forecasts that match or beat the full-input model at most 1–6 month lead times, while reducing i

desk verdict Incremental workshop paper with a clear pipeline but an overstated central claim; the causal-discovery leakage question is the one thing a referee must pin down. read the letter →

arxiv 2509.09128 v1 pith:SBO6WT5C submitted 2025-09-11 cs.LG

Emam Hossain , Md Osman Gani This is my paper
classification cs.LG
keywords ArcticseaiceextentcausaldiscoveryMVGCPCMCI+GRU-LSTMfeatureselectiontimeseriesforecastinginterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that letting causal discovery pick the inputs, rather than feeding a deep network every available variable, makes Arctic sea-ice-extent forecasts better and easier to interpret. It applies MVGC and PCMCI+ to 43 years of daily and monthly ocean-atmospheric data, then trains a GRU-LSTM on the selected features for 1- to 6-month lead times. The reported tables show the causal-input models generally matching or beating the full-input model on most horizons, with fewer inputs, although no single variant wins everywhere. If the claim holds, the practical payoff is cheaper, more trustworthy Arctic sea-ice forecasting and a template for causality-aware deep learning in other high-dimensional time-series problems.

What carries the argument

The pipeline couples two causal-discovery algorithms—MVGC, which tests whether lagged multivariate series improve prediction of the target, and PCMCI+, which builds lagged causal graphs via momentary conditional independence—with a GRU-LSTM whose 21-timestep look-back mirrors the max lag used in discovery. The causal graph is the mechanism that cuts the input set from ten ocean-atmospheric variables to a handful (daily: longwave radiation, snowfall, sea surface salinity, surface pressure; monthly: longwave radiation, SST), and the recurrent network trains only on those nodes. It is what lets the authors claim the model learns drivers rather than correlated noise.

What would settle it

Run the same GRU-LSTM training but restrict causal discovery to data through 2013 only. If the selected feature sets or forecast errors change materially across lead times, the paper's central claim is partly an artifact of leakage; if they stay the same and causal models still win, the claim survives.

Watch

Extended reading notes

Core claim

The central discovery is that a small set of causally selected predictors can carry most of the predictive information for Arctic SIE. MVGC flags all variables except sea surface temperature as Granger-causal drivers; PCMCI+ narrows the daily input set to longwave radiation, snowfall, sea surface salinity, surface pressure, and SIE itself, and the monthly set to longwave radiation, SST, and SIE. Trained in a GRU-LSTM with a 21-step look-back, these reduced sets produce errors and R2 values that, across the 1–6 month horizons, are usually on par with or better than the full 10-variable model—the exact winning variant varies by lead time and resolution. The paper also finds that transferring d

Load-bearing premise

The paper's central result depends on causal feature selection being performed on data that ends before the test window (2014 onward), a constraint it never explicitly states; if the discovery algorithms saw the test years, the reported accuracy gains would be inflated by target leakage.

Editorial extensions

If this is right

  • If causal feature selection works as claimed, operational Arctic sea-ice models can drop redundant inputs, cutting training and inference cost without losing forecast skill.
  • Daily-resolution causal features, transferred to monthly forecasts, give the best 5–6 month lead-time accuracy, suggesting that high-resolution causal structure can support seasonal predictions.
  • The causal graphs themselves are an interpretability deliverable: they name which atmospheric and oceanic variables directly drive SIE, so forecasts can be explained by their drivers.
  • The framework is proposed as a template for other high-dimensional dynamical systems where correlation-based deep learning is prone to overfitting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A reader should treat the headline claim as horizon-dependent: the paper's own tables show the full-input model winning at 1-month daily RMSE and several other cells, so the defensible claim is 'causal subsets usually match or beat full inputs,' not 'always win.'
  • The paper's Section 4.4 and Algorithm 1 never state whether causal discovery was restricted to data through 2013. If MVGC and PCMCI+ saw the 2014–2021 test years, the reported gains would be inflated by target leakage; a direct check would be to rerun discovery on pre-2014 data only.
  • The repurposing of daily causal features for monthly forecasts is an idea worth testing beyond SIE: if high-resolution causal structure transfers across temporal aggregations, it gives a cheap way to improve coarse-resolution forecasts.
  • The causal graphs are not evaluated against physical knowledge; a plausible extension would check whether the selected drivers (e.g., radiation and snowfall) match known sea-ice mechanisms and whether the discovered lags are physically sensible.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a causality-aware deep learning framework for Arctic sea ice extent (SIE) forecasting. It combines causal feature selection methods (MVGC and PCMCI+) with a hybrid GRU-LSTM model. The authors use daily and monthly ocean-atmospheric data from 1979–2021, select subsets of predictors via the two causal discovery algorithms, and compare models trained on the full feature set (DL_vanilla) against models trained on causally selected features (DL_GC, DL_PCMCI+, and DL_DPCMCI+ for monthly predictions). The main claim is that incorporating causal inputs improves prediction accuracy and interpretability across lead times from 1 to 6 months. The paper reports RMSE, MAE, and R² metrics for daily and monthly models and includes Algorithm 1 outlining the pipeline.

Significance. If the central claim were validated, the paper would provide a practical demonstration that causal feature selection can reduce input dimensionality and improve SIE forecasting, with potential applicability to other dynamical systems. The authors address a relevant and growing area—combining causal discovery with deep learning for climate prediction—and the framework is clearly described. The paper also promises public code and data. However, the empirical support is currently undermined by (i) a likely leakage path in the causal discovery step, (ii) an unfair feature comparison in which causal models additionally receive SIE as an input, and (iii) inconsistent wins across lead times that contradict the abstract's claim of improved accuracy 'across varying lead times.' These issues are fixable but require re-analysis.

major comments (4)
  1. [Section 4.4, Algorithm 1 Step 2] The manuscript states that GRU-LSTM models were trained using data up to 2013 and tested on 2014–2018 (daily) and 2014–2021 (monthly), but it never states that MVGC and PCMCI+ causal discovery was restricted to data before 2014. Algorithm 1 lists Step 2 (causal feature identification) before Step 3 (model training) but specifies no temporal cutoff for Step 2. If discovery used the full 1979–2021 record, the selected features in Table 2 and their lags were informed by SIE values from the exact test years, which is test-label leakage at the feature-selection stage. This would invalidate the attribution of any accuracy gain to causal selection. The authors must either explicitly confirm that discovery was run only on pre-2013 data or re-run the entire pipeline with an enforced temporal split and report whether the findings change.
  2. [Section 4.2, Table 2] Table 2 shows that all causal variants (DL_GC, DL_PCMCI+, DL_DPCMCI+) include SIE itself as an input feature, whereas DL_vanilla is listed as using 'All 10 ocean-atmospheric variables' and therefore excludes SIE. This is not an apples-to-apples comparison: causal models are given lagged SIE as an autoregressive input, which typically provides a strong advantage in sea ice forecasting. The reported improvements of DL_GC/DL_PCMCI+ over DL_vanilla could be due to the presence of the target's own history rather than to causal feature selection. The authors should either train a vanilla model that also includes lagged SIE as an input, or remove SIE from the causal feature sets, and then compare.
  3. [Section 5, Tables 3 and 4] The abstract's claim of 'improved prediction accuracy ... across varying lead times' is not supported by Tables 3 and 4. In Table 3, DL_vanilla achieves the lowest RMSE at the 1-month daily lead time (7.777 vs. 8.017 for DL_GC and 8.043 for DL_PCMCI+); causal-model wins are scattered across lead times (DL_GC at 2, 4, 5, and 6 months, DL_PCMCI+ at 1-month MAE and 3-month MAE). In Table 4, winners vary by lead time and metric, with DL_DPCMCI+ winning some RMSE rows but not others. Section 5 itself acknowledges 'no single model consistently outperforms the others across all forecast horizons.' This internal inconsistency needs to be resolved: either the central claim should be narrowed to specific configurations and lead times, or additional evidence should show a consistent advantage after fixing the leakage and feature-fairness issues.
  4. [Section 5, Tables 3 and 4] The evaluation lacks uncertainty quantification and a standard persistence baseline. The test periods are short (5 years daily, 8 years monthly), and no error bars or confidence intervals are reported across random seeds. Many RMSE differences are small (e.g., daily 1-month: 7.777 vs. 8.017, about 3%), so the observed pattern of wins and losses may be noise. In addition, persistence or climatology baselines are standard in SIE forecasting and would help assess whether any model has meaningful skill beyond trivial extrapolation. Please add repeated-seed statistics, statistical significance tests, and a persistence benchmark.
minor comments (6)
  1. [Table 2] Names are inconsistent: 'DL_vanilla' in Table 3 vs. 'DLvanilla' in Table 2, and similarly for other variants. Please standardize.
  2. [Section 4.3, Algorithm 1] Please clarify whether the 21-step lookback and maximum lag τ refer to 21 days, 21 months, or the same number of timesteps for both daily and monthly datasets.
  3. [Section 5, Figures 5 and 6] R² values are only presented visually; numeric R² values and error bars should be included in the tables or an appendix.
  4. [Section 4.1] The paper says 'ten ocean-atmospheric predictors and sea ice extent values' but Table 1 lists Sea Ice Extent as one of the variables. Clarify that SIE is the target and not used as a predictor in the vanilla model.
  5. [Section 5] The daily test period is 2014–2018, but the daily data are described as spanning 1979 to 2018. Please explain how multi-month lead times are generated from daily data and whether predictions are aggregated to monthly values.
  6. [General] The GitHub URL in the footer ('learning what matters') contains spaces; ensure the repository link is valid and points to the exact code used for the experiments.

Circularity Check

1 steps flagged · score 4.0 of 10

Potential test-set leakage in causal feature selection means the reported 'prediction' gains may partly reduce to the selector's fit to the test labels; no explicit circular derivation elsewhere.

  1. fitted input called prediction [Algorithm 1 Step 2 and Section 4.4]
    "Step 2: Causal Feature Identification ... Use MVGC to extract causal variables C_GC ⊆ X_t. Apply PCMCI+ to determine causal subsets C_PCMCI+ ⊆ X_t for both temporal resolutions. ... The GRU-LSTM models were trained using historical data up to the year 2013, reserving 10% of the training split for validation. The test set comprised observations from 2014–2018 for the daily dataset and 2014–2021 for the monthly dataset."

    Algorithm 1 takes as input the full multivariate series D (1979–2021 monthly / 1979–2018 daily, Section 4.1) and does not restrict Step 2 to data through 2013. Section 4.4 restricts only model training to pre-2014, not the MVGC/PCMCI+ discovery. As written, the causal feature sets in Table 2 can be estimated using SIE values from the exact years later used as test labels. The 'causal inputs' are then fed to the GRU-LSTM and the resulting RMSE/MAE/R² on 2014–2021 are reported as improvements. If discovery saw the test period, the comparison is a feature selector fit to the outcome, then renamed 'prediction' — the reported gains are not independent of the selection step.

full rationale

The paper contains no mathematical derivation that is circular by construction: MVGC/PCMCI+ are standard causal-discovery tools, the GRU-LSTM is a standard sequence model, and the empirical comparison against DL_vanilla is a legitimate experimental design. However, the central claim ('incorporating causal inputs leads to improved prediction accuracy... across varying lead times') rests on the causal features having been chosen without information from the evaluation period. The paper never states that MVGC/PCMCI+ were run only on data through 2013, and Algorithm 1's Step 2 precedes the train/test split without specifying a temporal cutoff. This creates a concrete leakage path: features selected with full-record SIE (including 2014–2018/2021) can make the test scores an artifact of the selector's fit to those labels. The paper also undercuts its own headline by noting 'no single model consistently outperforms the others across all forecast horizons,' so the claimed uniform improvement is not actually shown. Self-citations to the authors' earlier sea-ice/causality papers are present but not load-bearing for the central empirical comparison, since the underlying methods are cited to Barnett/Seth and Runge. Because the leakage is a potential reduction of the prediction claim to the selection step rather than a demonstrated equation-level identity, a moderate score is appropriate.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central comparison rests on unspecified discovery thresholds, a fixed lag window, and several untested causal assumptions (causal sufficiency, stationarity, resolution invariance). These are assumptions the reader pays for upstream rather than results the paper establishes.

free parameters (4)
  • Maximum lag tau = 21
    Used as the look-back window and max lag for MVGC/PCMCI+; the paper gives no data-driven justification or sensitivity analysis (Section 4.3, Algorithm 1).
  • GRU/LSTM hidden sizes and dropout = 64/128/64 units, 20% dropout
    Chosen without ablation; no sensitivity reported (Section 4.3).
  • MVGC/PCMCI+ significance thresholds = not reported
    The causal subsets in Table 2 depend on unspecified test thresholds; without them the feature sets cannot be reproduced (Section 4.2).
  • Train/validation split fraction = 90/10
    Chosen without discussion; minor but affects early stopping behavior (Section 4.4).
assumptions (4)
  • domain assumption Causal sufficiency: no unobserved confounders between the 10 predictors and SIE
    MVGC and PCMCI+ infer causal structure from observed data; the paper does not discuss unmeasured drivers (e.g., oceanic heat transport, clouds) that could confound links (Section 4.2).
  • domain assumption Stationarity of causal structure across 1979 to 2021
    The causal graph is estimated once and assumed to hold for the test years; Arctic regime change makes this questionable (Sections 4.2 and 4.4).
  • standard math PCMCI+ conditional independence tests are correctly specified for autocorrelated nonlinear data
    This is the method's published guarantee, but the paper does not verify assumptions such as faithfulness or absence of contemporaneous hidden confounders (Section 3.1).
  • ad hoc to paper Daily causal features are valid for monthly prediction
    DL_DPCMCI+ uses daily-discovered features for monthly forecasting; this transfer assumes resolution invariance (Section 4.4, Table 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning What Matters: Causal Time Series Modeling for Arctic Sea Ice Prediction." pith.science (2026). https://pith.science/paper/SBO6WT5C

@misc{pith2026250909128,
  author       = {Pith},
  title        = {Pith review of: Learning What Matters: Causal Time Series Modeling for Arctic Sea Ice Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SBO6WT5C}},
  note         = {Machine review of arXiv:2509.09128}
}
read the original abstract

Conventional machine learning and deep learning models typically rely on correlation-based learning, which often fails to distinguish genuine causal relationships from spurious associations, limiting their robustness, interpretability, and ability to generalize. To overcome these limitations, we introduce a causality-aware deep learning framework that integrates Multivariate Granger Causality (MVGC) and PCMCI+ for causal feature selection within a hybrid neural architecture. Leveraging 43 years (1979-2021) of Arctic Sea Ice Extent (SIE) data and associated ocean-atmospheric variables at daily and monthly resolutions, the proposed method identifies causally influential predictors, prioritizes direct causes of SIE dynamics, reduces unnecessary features, and enhances computational efficiency. Experimental results show that incorporating causal inputs leads to improved prediction accuracy and interpretability across varying lead times. While demonstrated on Arctic SIE forecasting, the framework is broadly applicable to other dynamic, high-dimensional domains, offering a scalable approach that advances both the theoretical foundations and practical performance of causality-informed predictive modeling.

Figures

Figures reproduced from arXiv: 2509.09128 by the authors.

Figure 1
Figure 1. Approximately 38% decline in Arctic September sea ice, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of how causal discovery algorithms identify [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed causal deep learning framework for Arctic Sea Ice Extent (SIE) prediction. Daily and monthly temporal [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Causal graphs generated by PCMCI+ for (a) daily and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: R 2 values for daily models across the lead times. PCMCI+-identified features for near-term prediction tasks. For intermediate lead times such as 2 months, DLDPCMCI+ outperforms all other models. This model transfers causal features identified from daily data into the …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 4 linked inside Pith

  1. [1]

    Sea ice forecasting using attention-based ensemble lstm.arXiv preprint arXiv:2108.00853,

    [Aliet al., 2021 ] Sahara Ali, Yiyi Huang, Xin Huang, and Jianwu Wang. Sea ice forecasting using attention-based ensemble lstm.arXiv preprint arXiv:2108.00853,

    arXiv 2021

  2. [3]

    The mvgc multivariate granger causality toolbox: a new approach to granger-causal inference.Journal of neuro- science methods, 223:50–68,

    [Barnett and Seth, 2014] Lionel Barnett and Anil K Seth. The mvgc multivariate granger causality toolbox: a new approach to granger-causal inference.Journal of neuro- science methods, 223:50–68,

    2014

  3. [6]

    Comparative analysis of recurrent neural networks in stock price prediction for different frequency domains.Algorithms, 14(8):251,

    [Deyet al., 2021 ] Polash Dey, Emam Hossain, Md Ishti- aque Hossain, Mohammed Armanuzzaman Chowdhury, Md Shariful Alam, Mohammad Shahadat Hossain, and Karl Andersson. Comparative analysis of recurrent neural networks in stock price prediction for different frequency domains.Algorithms, 14(8):251,

    2021

  4. [7]

    Data-driven emulation of melt ponds on arctic sea ice.EGUsphere, 2024:1–18,

    [Driscollet al., 2024 ] Simon Driscoll, Alberto Carrassi, Julien Brajard, Laurent Bertino, Einar ´Olason, Marc Boc- quet, and Amos Lawless. Data-driven emulation of melt ponds on arctic sea ice.EGUsphere, 2024:1–18,

    2024

  5. [10]

    Investigating causal rela- tions by econometric models and cross-spectral methods

    [Granger, 1969] Clive WJ Granger. Investigating causal rela- tions by econometric models and cross-spectral methods. Econometrica: journal of the Econometric Society, pages 424–438,

    1969

  6. [14]

    Correlation to causation: A causal deep learn- ing framework for arctic sea ice prediction

    [Hossainet al., 2025 ] Emam Hossain, Muhammad Hasan Ferdous, Jianwu Wang, Aneesh Subramanian, and Md Os- man Gani. Correlation to causation: A causal deep learn- ing framework for arctic sea ice prediction. In2025 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), pages 62–67. IEEE,

    2025

  7. [15]

    Foreign exchange currency rate prediction us- ing a gru-lstm hybrid network.Soft Computing Letters, 3:100009,

    [Islam and Hossain, 2021] Md Saiful Islam and Emam Hos- sain. Foreign exchange currency rate prediction us- ing a gru-lstm hybrid network.Soft Computing Letters, 3:100009,

    2021

  8. [16]

    [Kimet al., 2025 ] Young Jun Kim, Hyun-cheol Kim, Dae- hyeon Han, Julienne Stroeve, and Jungho Im. Long- term prediction of arctic sea ice concentrations using deep learning: Effects of surface temperature, radiation, and wind conditions.Remote Sensing of Environment, 318:114568,

    2025

Show all 25 references
  1. [17]

    Hierarchical information-sharing convolutional neural network for the prediction of arctic sea ice concentration and velocity.IEEE Transactions on Geoscience and Remote Sensing,

    [Koo and Rahnemoonfar, 2024] Younghyun Koo and Maryam Rahnemoonfar. Hierarchical information-sharing convolutional neural network for the prediction of arctic sea ice concentration and velocity.IEEE Transactions on Geoscience and Remote Sensing,

    2024

  2. [18]

    Advancing arctic sea ice remote sensing with ai and deep learning: now and future.EGUsphere, 2024:1– 36,

    [Liet al., 2024 ] Wenwen Li, Chia-Yu Hsu, and Marco Tedesco. Advancing arctic sea ice remote sensing with ai and deep learning: now and future.EGUsphere, 2024:1– 36,

    2024

  3. [19]

    Physics-informed deep convolutional network for com- bined sea ice concentration and velocity prediction.Ocean Engineering, 313:119440,

    [Liuet al., 2024 ] Quanhong Liu, Yangjun Wang, Ren Zhang, Lujun Zhang, Hengqian Yan, and Kefeng Liu. Physics-informed deep convolutional network for com- bined sea ice concentration and velocity prediction.Ocean Engineering, 313:119440,

    2024

  4. [20]

    Causality- inspired models for financial time series forecasting.arXiv preprint arXiv:2408.09960,

    [Oliveiraet al., 2024 ] Daniel Cunha Oliveira, Yutong Lu, Xi Lin, Mihai Cucuringu, and Andre Fujita. Causality- inspired models for financial time series forecasting.arXiv preprint arXiv:2408.09960,

    2024 arXiv

  5. [21]

    Basic books,

    [Pearl and Mackenzie, 2018] Judea Pearl and Dana Macken- zie.The book of why: the new science of cause and effect. Basic books,

    2018

  6. [23]

    Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets

    [Runge, 2020] Jakob Runge. Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets. InConference on Uncertainty in Ar- tificial Intelligence, pages 1388–1397. Pmlr,

    2020

  7. [25]

    Stdnet: Spatio- temporal decompose network for predicting arctic sea ice concentration.Remote Sensing, 16(23):4534, 2024

    [Zhuet al., 2024 ] Xu Zhu, Jing Wang, Guojun Wang, Yang- ming Jiang, Yi Sun, and Huihui Zhao. Stdnet: Spatio- temporal decompose network for predicting arctic sea ice concentration.Remote Sensing, 16(23):4534, 2024

    2024

  8. [1969]

    A survey on causal discovery meth- ods for iid and time series data.Transactions on Machine Learning Research,

    [Hasanet al., 2023 ] Uzma Hasan, Emam Hossain, and Md Osman Gani. A survey on causal discovery meth- ods for iid and time series data.Transactions on Machine Learning Research,

    2023

  9. [1996]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation.arXiv preprint arXiv:1406.1078,

    [Choet al., 2014 ] Kyunghyun Cho, Bart Van Merri ¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation.arXiv preprint arXiv:1406.1078,

    2014 arXiv

  10. [1997]

    A novel deep learning approach to predict air qual- ity index

    [Hossainet al., 2020 ] Emam Hossain, Mohd Arafath Uddin Shariff, Mohammad Shahadat Hossain, and Karl Anders- son. A novel deep learning approach to predict air qual- ity index. InProceedings of International Conference on Trends in Computational and Cognitive Engineering: Pro-...

    2020

  11. [2014]

    [Cavalieriet al., 1996 ] Donald Cavalieri, Claire Parkinson, Per Gloersen, and H. Zwally. Sea ice concentrations from nimbus-7 smmr and dmsp ssm/i-ssmis passive microwave data, version 1,

    1996

  12. [2018]

    Sicnet season v1

    [Renet al., 2024 ] Yibin Ren, Xiaofeng Li, and Yunhe Wang. Sicnet season v1. 0: a transformer-based deep learning model for seasonal arctic sea ice prediction by integrating sea ice thickness data.Geoscientific Model Development Discussions, 2024:1–20,

    2024

  13. [2020]

    Sifm: A foun- dation model for multi-granularity arctic sea ice forecast- ing.arXiv preprint arXiv:2410.14732,

    [Xuet al., 2024 ] Jingyi Xu, Yeqi Luo, Weidong Yang, Keyi Liu, Shengnan Wang, Ben Fei, and Lei Bai. Sifm: A foun- dation model for multi-granularity arctic sea ice forecast- ing.arXiv preprint arXiv:2410.14732,

    2024 arXiv

  14. [2021]

    Seasonal arctic sea ice forecasting with probabilistic deep learning.Nature communications, 12(1):5124,

    [Anderssonet al., 2021 ] Tom R Andersson, J Scott Hosking, Mar´ıa P ´erez-Ortiz, Brooks Paige, Andrew Elliott, Chris Russell, Stephen Law, Daniel C Jones, Jeremy Wilkinson, Tony Phillips, et al. Seasonal arctic sea ice forecasting with probabilistic deep learning.Nature commun...

    2021

  15. [2023]

    Long short-term memory

    [Hochreiter, 1997] S Hochreiter. Long short-term memory. Neural Computation MIT-Press,

    1997

  16. [2024]

    Subra- manian, Emam Hossain, Md Osman Gani, Alison F

    [Dunmireet al., 2025 ] Devon Dunmire, Aneesh C. Subra- manian, Emam Hossain, Md Osman Gani, Alison F. Ban- well, Hammad Younas, and Brendan Myers. Greenland ice sheet wide supraglacial lake evolution and dynamics: Insights from the 2018 and 2019 melt seasons.Earth and Space Sc...

    2025

  17. [2025]

    Timegraph: Synthetic benchmark datasets for robust time-series causal discov- ery

    [Ferdouset al., 2025 ] Muhammad Hasan Ferdous, Emam Hossain, and Md Osman Gani. Timegraph: Synthetic benchmark datasets for robust time-series causal discov- ery. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, pages 5425–5435,

    2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.