acceptodds
Under review as a conference paper at ICLR 2027

FORMA: Feedback-Conditioned Spatial Memory for Field Forecasting

Abstract

Long-horizon geophysical forecasting requires a predictive state that can remember persistent seasonal structure while rapidly adapting to moving fronts, changing spatial interactions, and unreliable observations. Existing forecasting architectures provide powerful temporal or spatial modeling mechanisms, but coupling long-term memory with observation-dependent spatial adaptation remains difficult. We introduce FORMA, a forecasting architecture centered on Field Memory, a field-valued predictive state whose evolution is regulated by feedback from the observed geophysical field. Seasonal phase, observation reliability, and interface dynamics determine what information is retained, what new evidence is written, and how memory propagates across space. FORMA implements this idea with an RWKV-style temporal recurrence together with efficient global and local spatial interaction, preserving spatial tokens throughout the historical scan while maintaining linear complexity in history length and grid size. On 90-day sea-surface-temperature forecasting, FORMA reduces area RMSE by 19.1% over a matched RWKV recurrence and by 7.9% over the strongest baseline in our comparison, with consistent gains in interface and basin-scale accuracy. The same architecture also transfers to Arctic sea-ice concentration forecasting, reducing concentration RMSE by 4.3% and sea-ice-extent error by 25.4%. These results demonstrate that explicitly modeling a feedback-regulated spatial predictive state is a promising alternative for long-horizon forecasting of evolving geophysical fields.

Primary area:
learning on time series and dynamical systems
Keywords:
State Space Models, Recurrent Neural Networks, Dynamical Systems, Geophysical Forecasting

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_HdHVOpenReview ↗
Rating
2
Confidence
4(+2)
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes FORMA, a spatiotemporal forecasting architecture designed for 90-day prediction of evolving geophysical fields, specifically sea-surface temperature (SST) and Arctic sea-ice concentration (SIC). Its main contribution is Field Memory, a recurrent spatial state module that combines an RWKV-style temporal scan with factorized global and local spatial interactions, achieving linear complexity in history length and grid size. To adapt state evolution to nonstationary dynamics, the memory's retention, writing, and spatial propagation operators are regulated by three causal feedback signals: seasonal phase encodings, a date-level observation-reliability descriptor, and learned spatial interface features . The terminal memory state is combined with the latest observation and mapped to all future horizons via a static multi-horizon decoder trained with field, interface, regional-mean, and temporal-difference losses. The authors validate FORMA through a synthetic translating-front diagnostic, 90-day global SST forecasting on NOAA OISSTv2.1, and Arctic SIC forecasting on NSIDC CDR v6, supplemented by ablations on the feedback signals

Critical strengths

Explicitly regulating temporal retention, evidence writing, and local neighbor routing via seasonal phase and spatial gradient/threshold interfaces is a clean conceptual design. It directly addresses the coexistence of slow seasonal modes and sharp moving boundaries in geophysical fields while maintaining bounded state propagation (∥Ft∥∞<1\|F_t\|_\infty < 1).

The same architecture transfers from global SST to Arctic SIC, achieving consistent improvements in both bulk field accuracy and front/ice-edge localization.

Critical weaknesses

Weaknesses

  1. The novelty boundary needs sharper positioning. Observation-dependent propagation of spatial recurrent states has precedents such as TrajGRU (Shi et al., 2017). The normalized factorization in Eq. (9) also follows the aggregation form used in Efficient Attention (Shen et al., 2021). FORMA’s contribution appears to lie in integrating these mechanisms with seasonal, reliability, and interface conditioning. The paper should explain more clearly what this integration enables beyond existing adaptive recurrent architectures.

  2. The comparisons do not sufficiently isolate the proposed memory mechanism. The matched RWKV control removes seasonal, reliability, and interface conditioning together with spatial propagation. Likewise, the no-interface variant also removes interface supervision. These comparisons establish the benefit of the complete system, but leave the architectural contribution entangled with additional information and training objectives. This matters because removing seasonal conditioning causes a particularly large degradation.

  3. Given the strong effect of seasonal conditioning, the current results leave uncertainty about how much improvement comes from predicting anomalous evolution and interface displacement. A focused evaluation of forecast skill beyond the seasonal background would substantially strengthen the central claim.

Questions for authors
  1. How much of the improvement remains when a strong recurrent baseline receives the same seasonal and reliability inputs and comparable training objectives? Can the interface supervision be retained while changing only the spatial update?

  2. Does FORMA retain its advantage over seasonal climatology and anomaly persistence, particularly for anomalous conditions and fronts or ice edges displaced from their usual seasonal positions?

  3. What is the exact construction of the reliability score qtq_t, and which quality variations in the original observation products explain its substantial benefit on unperturbed histories?

11 comments
AuthorsOpenReview ↗

Response to Reviewer HdHV

Dear Reviewer HdHV,

We sincerely appreciate your time and effort in reviewing our paper. Thank you for your positive assessment of FORMA’s conceptual design, particularly recognizing that explicitly regulating retention, evidence writing, and local routing through seasonal phase and interface features addresses the coexistence of slow seasonal modes and sharp moving boundaries. We also appreciate your recognition that this design maintains bounded state propagation. We are glad that you acknowledged the transfer of the same architecture from global SST to Arctic SIC, with consistent improvements in field accuracy and front/ice-edge localization. We hope our responses below address your concerns. We would greatly appreciate it if you would consider updating your score in light of these clarifications and results.

W1. The novelty and advantages of Field Memory.

A1. Our contribution is a field-valued predictive state whose retention, writing, and propagation jointly respond to seasonal phase, reliability, and interfaces. This coupling lets persistent context survive while new evidence updates moving boundaries, and lets stored evidence interact at both local and basin scales throughout the historical scan. This unified memory has linear cost in history length and grid size at fixed width and rank.

TrajGRU [1] adapts recurrent connections using the input and previous state. FORMA couples temporal evidence accumulation with explicit retention, writing, and complementary global/local transport, conditioned on the observed field and its context. Efficient Attention [2] aggregates values projected from current features; FORMA uses its factorization to transport accumulated memory under new observations.

Eq. (11) expresses how each dated write contributes through subsequent transitions; Appendix C bounds propagation with factors and writes fixed. Table 4(a) supports the conditioning components, and SST/SIC results show the design's forecasting value. This explanation will be included in the revised manuscript.

W2. Additional ablations to isolate the memory mechanism.

A2. Thank you for this question. We evaluate the complete coupled mechanism under matched inputs and supervision, then examine spatial exchange within it.

The original ablations examine components within this design. We additionally replace the complete Field Memory block with ConvLSTM. Both receive identical field/validity, phase, q, and interface features through the same input adapter, retaining the encoder, decoder, interface supervision, five losses and weights, data splits, and checkpoint-selection rule. Each model is trained separately.

The temporal-only control retains the encoder, temporal accumulators, input mixing, retention and write gates, and decoder. It removes only the memory-transport term while keeping the retention coefficient fixed.

Table R1. Matched information and supervision. Errors in 10−2∘C10^{-2}{}^\circ\mathrm{C}.

Memory mechanismArea RMSEInterface RMSEBasin error
Conditioned temporal-only108.7127.46.2
Conditioned ConvLSTM104.9121.65.9
FORMA with self routes106.8124.16.1
Full Field Memory93.9109.65.5

The 10.5% area-RMSE reduction over ConvLSTM evaluates the complete mechanism. The self-route variant separately replaces global/local operators by their row-sum diagonals, retaining retention, writing, input mixing, and supervision. Its 12.1% gap tests spatial exchange within Field Memory. Interface heads, targets, and losses remain active in both comparisons.

References

[1] Shi et al. TrajGRU. NeurIPS 2017. https://arxiv.org/abs/1706.03458

[2] Shen et al. Efficient Attention. WACV 2021. https://arxiv.org/abs/1812.01243

AuthorsOpenReview ↗

Response to Reviewer HdHV (Continued)

W3. Does the phase gain reflect skill beyond the seasonal background?

A3. We understand why the large phase effect raises this question. Phase regulates all three operations of Field Memory. Table 4(a) examines this coupled design: removing phase, reliability, or the interface components raises area RMSE from 93.9 to 153.9, 120.7, or 112.8, respectively. Each contributes to the performance of the full architecture; the no-interface ablation includes its associated supervision.

This ablation measures phase within FORMA; comparing the ablated model with another architecture cannot isolate the latter's benefit from phase. We therefore project and broadcast the same five phase features into each baseline, preserving its backbone, losses, and training protocol. Without/with phase, area RMSE is 102.0/101.8 (3D U-Net), 103.4/103.1 (Earthformer), 117.0/116.5 (STDMamba), and 153.9/93.9 (FORMA), in 10−2∘C10^{-2}{}^\circ\mathrm{C}.

These controls highlight FORMA's effective use of seasonal context. Our extensive experiments and related studies [3,4] offer clues about architecture and conditioning, without fully explaining the gain's magnitude or when comparable gains arise. We report this observation and the controls faithfully.

We also appreciate your suggestion to evaluate departures from the seasonal background. After subtracting training-only seasonal means, the day-90 anomaly correlation is .390 for FORMA versus .345 for 3D U-Net. On displaced SST fronts, boundary distance is 98.6 versus 109.7 km. These measures directly assess anomalous evolution and interface location, complementing the evidence for the integrated design. This explanation and evaluation will be included in the revised manuscript.

Q1. What remains with matched inputs and losses? Can interface supervision be retained?

A1. Yes. Table R1 retains identical inputs and all five losses: the area-RMSE gains are 10.5% over conditioned ConvLSTM and 12.1% over self routes. The latter changes only memory propagation, preserving both interface heads, targets, and loss weights. Its transport coefficient remains 0.1 and route weights 1/2, with each operator's row sum preserved.

Q2. Does the advantage extend beyond climatology and anomaly persistence, especially for unusual interfaces?

A2. Our experiments prioritize modern forecasting architectures because our contribution is a new memory mechanism. This is a deliberate research choice. We take issue with the emphasis on these elementary references: it gives insufficient attention to the methodological work and comparisons already presented. We provide the requested comparisons below, while maintaining that they supplement the evaluation of our architectural contribution.

We subtract training-only (1982–2018) location-specific day-of-year means from predictions and observations. Anomaly correlation is the area-weighted normalized inner product of predicted and observed anomaly fields, averaged over test origins.

Table R2. Lead-day SST. RMSE: 10−2∘C10^{-2}{}^\circ\mathrm{C}. Table 1's 93.9 covers the full 90-day window.

Lead dayFORMA RMSEFORMA anomaly correlation3D U-Net anomaly correlation
3087.5.694.670
6099.1.572.529
90111.6.390.345

With SSb=1−MSEFORMA/MSEbSS_b=1-\mathrm{MSE}_{FORMA}/\mathrm{MSE}_b, the skill score against seasonal climatology is .429/.275/.092 at days 30/60/90; against anomaly persistence, −.030/.081/.095. Anomaly persistence performs slightly better at day 30, while FORMA achieves lower MSE at days 60 and 90. Dates, masks, and weights are identical.

References

[3] Zeng et al. Are Transformers Effective for Time Series Forecasting? AAAI 2023. https://arxiv.org/abs/2205.13504

[4] Wang et al. TimeXer. NeurIPS 2024. https://arxiv.org/abs/2402.19072

AuthorsOpenReview ↗

Response to Reviewer HdHV (Final Part)

A2 (continued). Event subsets.

Day-90 event subsets use training-period, season-specific 90th-percentile thresholds for area-weighted anomaly magnitude or displacement from the seasonal boundary. All models share cases; test observations define retrospective groups.

Table R3. Event subsets. Units: SST errors, 10−2∘C10^{-2}{}^\circ\mathrm{C}; IIEE, 10−210^{-2} million km²; symmetric boundary distance, km. n counts forecast origins.

Subset (n)Metric(s)Baseline modelBaseline valueFORMA
Strong SST anomalies (126)Anomaly RMSE / correlation3D U-Net148.6 / .682137.4 / .731
Displaced SST fronts (112)Interface RMSE / distance3D U-Net164.8 / 109.7151.2 / 98.6
Displaced SIC edges (118)IIEE / distanceIceMamba143.8 / 79.1137.9 / 74.6

Relative to the neural forecasting baselines in Table R3, FORMA reduces the first-listed error by 7.5%, 8.3%, and 4.1% on the strong-SST-anomaly, displaced-SST-front, and displaced-SIC-edge subsets, respectively. Symmetric boundary distance decreases by 10.1% on displaced SST fronts and 5.7% on displaced SIC edges.

We also compare against seasonal climatology and anomaly persistence on the same subsets. In the same subset order, the first-listed error decreases by 24.6%, 18.4%, and 14.0% relative to seasonal climatology, and by 7.8%, 6.7%, and 5.0% relative to anomaly persistence.

Q3. How is q constructed, and which original-data quality variations explain its benefit?

A3. Thank you for asking how the score explains the clean-history result. It is computed as qt=0.55vt+0.25e−δt/0.10+0.20,q_t=0.55v_t+0.25e^{-\delta_t/0.10}+0.20, where vtv_t is the pre-imputation valid fraction within the supplied mask, or full grid if none is supplied. Coverage and finite-value weights combine into 0.55; 0.20 includes the default metadata contribution. δt\delta_t is the mean absolute change over jointly valid cells in consecutive normalized fields. SST uses training-only 0.5/99.5 percentile bounds, clipped to [0,1]; SIC uses [0,1]. At the first dataset date or without valid pairs, δt=0\delta_t=0.

“Clean history” means no artificial corruption was applied. With full valid coverage, qt=0.75+0.25e−δt/0.10q_t=0.75+0.25e^{-\delta_t/0.10}: temporal consistency still varies. This signal summarizes the changing input evidence, including real dynamics. OISST also has interpolation uncertainty and evolving observational inputs [5], but those product properties are not separately identified by this scalar.

The no-trust variant fixes q at one throughout temporal decay, retention, writing, and global routing, and removes its write multiplier. Thus, 120.7 versus 93.9 tests conditioning on day-to-day consistency. The same q change can accompany real field evolution or an observation artifact; this ablation does not distinguish those causes. It supports using this summary of incoming evidence to regulate memory updates. This explanation will be included in Appendix B.3.

Thank you for your careful review and constructive suggestions. We will further clarify the contribution of Field Memory and detail the matched ablations, forecast-skill evaluation, and reliability-score construction in the revised manuscript and supplementary materials. We will also release the source code to facilitate reproduction of the results and further research.

References

[5] NOAA/NCEI. OISST. https://www.ncei.noaa.gov/products/optimum-interpolation-sst

Reviewer_7GHqOpenReview ↗
Rating
1(−1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper proposes FORMA, a spatiotemporal forecasting architecture centered on Field Memory. The model combines RWKV-style temporal aggregation with low-rank global and interface-conditioned local spatial propagation. Seasonal phase, observation reliability, and interface features condition memory retention, evidence writing, and spatial interactions. Experiments on sea-surface temperature and Arctic sea-ice concentration forecasting show improvements over the evaluated baselines.

Critical strengths
  1. The paper addresses a relevant forecasting challenge: preserving persistent temporal information while adapting to evolving spatial boundaries and varying observation quality.
  2. The architecture integrates temporal aggregation and complementary global/local spatial interactions while maintaining linear complexity in history length and grid size at fixed feature width and route rank. Evaluation covers two geophysical fields and includes field accuracy, interface-related metrics, input perturbations, and history-length studies.
Critical weaknesses
  1. The interaction between the three conditioning signals is insufficiently examined. Seasonal context, observation reliability, and interface dynamics are related rather than independent. In particular, an apparent boundary change may reflect either genuine movement that requires rapid adaptation or observation noise that should be suppressed. Although the signals jointly condition state updates, the paper does not establish whether the proposed mechanism can distinguish these cases and produce appropriate retention and writing behavior.

  2. The framework presentation is difficult to follow. The correspondence between the architectural diagram and the mathematical definitions is not sufficiently explicit. In particular, the roles and connections of the temporal accumulators, spatial memory, interface-conditioned input mixing, and local memory propagation require clearer explanation.

  3. The reliability signal is insufficiently specified. Appendix B.3 lists valid coverage, finite-value status, temporal consistency, and product metadata, but does not provide a reproducible procedure for computing and combining them into q. Moreover, q is a date-level scalar broadcast across the spatial grid, which cannot explicitly represent regional differences in observation quality.

  4. The ablations do not adequately separate the benefit of auxiliary information from that of the proposed memory mechanism. Removing seasonal conditioning substantially increases SST area RMSE from 0.939 to approximately 1.539. However, the matched RWKV control also removes seasonal and other conditioning signals. Consequently, it remains unclear how much improvement comes from providing explicit calendar information versus the proposed way of updating memory.

  5. Statistical uncertainty is not reported. The experimental results do not include standard deviations or confidence intervals across repeated training runs. This limits confidence in the reproducibility of the improvements, particularly for small differences such as the SIC IIEE results.

Questions for authors

When genuine interface movement and observation noise produce similar temporal changes, how does FORMA distinguish them? Can the authors provide a controlled experiment showing the resulting reliability estimates and memory-update behavior? Can the authors specify the exact computation of q, including its component scores and normalization? How would the method handle observations that are reliable in some regions but corrupted in others? How does FORMA compare with a matched RWKV or conventional recurrent baseline supplied with the same seasonal, reliability, and interface information?

10 comments
AuthorsOpenReview ↗

Response to Reviewer 7GHq

Dear Reviewer 7GHq,

We sincerely appreciate your time and effort in reviewing our paper. Thank you for your positive assessment of the significance of our work, particularly recognizing the importance of preserving persistent temporal information while adapting to evolving spatial boundaries and varying observation quality. We also appreciate your recognition of FORMA’s integration of temporal aggregation with complementary global and local spatial interactions, while maintaining linear complexity in history length and grid size at fixed feature width and route rank. We are glad that you acknowledged our evaluation across two geophysical fields, covering field accuracy, interface-related metrics, input perturbations, and history-length studies. We hope our responses below address your concerns. We would greatly appreciate it if you would consider updating your score in light of these clarifications and results.

W1. Joint conditioning under genuine motion and observation noise.

A1. Thank you for highlighting this ambiguity. FORMA combines seasonal context, temporal consistency, and interface structure in its memory update. With the default reliability exponent of one, Γt=qt⋅σ(WΓ[U~t,Bt,pt,qt]+bΓ).\Gamma_t=q_t\cdot\sigma(W_\Gamma[\widetilde U_t,B_t,p_t,q_t]+b_\Gamma). Here U~t\widetilde U_t is the spatially mixed observation, BtB_t encodes its underlying representation and temporal change, and ptp_t gives phase. q scales evidence acceptance, while the learned gate varies across locations and channels. Retention uses [U~t,pt,qt][\widetilde U_t,p_t,q_t]; local routing uses [Bt,pt][B_t,p_t]. Interface-conditioned input mixing also brings boundary information into retention. Thus, similar temporal changes can receive different updates through their spatial structure and context.

The submitted front and perturbation experiments examine motion and corruption separately. To test their distinction directly, we specify paired histories in Appendix D.1's translating-front setting: coherent displacement versus spatially incoherent corruption, matched in phase, coverage, and mean absolute innovation. Each target follows its underlying clean dynamics. Holding q fixed tests the learned response to spatial structure.

At full coverage and δt=0.04\delta_t=0.04, both cases give qt=0.9176q_t=0.9176. We measure the frozen model's gates over the same predefined boundary band and channels, and RMSE against each clean future in normalized field units. Table D1. Controlled diagnostic. q is calculated. D and Γ are band/channel means; Γ includes the q multiplier. A uniform-write control replaces Γ within the evaluation band by its spatial mean per channel at each step, retaining the same model and q.

Input changeqDΓFull RMSEUniform-write RMSE
Coherent motion.9176.742.386.046.058
Matched corruption.9176.891.164.061.077

In this diagnostic, coherent motion receives more writing and lower retention; corruption receives less writing and higher retention. Full-model RMSE is lower than the uniform-write control in both cases. Together these observations illustrate a spatially selective response beyond the shared q. They do not make q a motion/noise classifier. This explanation will be included in the revised methods.

W2. Correspondence between architecture and equations.

A2. Thank you for identifying this presentation issue. The key distinction is temporal evidence accumulation versus spatial memory propagation:

ComponentOperation and role
Input mixing, Eq. (3)Mix current observation UtU_t with neighbors to obtain U~t\widetilde U_t.
Accumulators, Eqs. (4–6)St,NtS_t,N_t accumulate weighted values and normalization; their readout gives EtE_t.
Spatial memory, Eqs. (7–10)Retain and transport Ht−1H_{t-1}, then add Γt⊙Et\Gamma_t\odot E_t.
DecoderCombine terminal memory and latest observation features to predict all leads.

Input mixing and local memory propagation share routing weights but act on current features and previous memory, respectively. The write gate connects the accumulators' evidence to memory. We will include this mapping in the methods and figure caption, labeling each state and connection.

AuthorsOpenReview ↗

Response to Reviewer 7GHq (Continued)

W3. Reproducible definition of q and regional quality variation.

A3. The field-derived score is qt=0.55vt+0.25e−δt/0.10+0.20,q_t=0.55v_t+0.25e^{-\delta_t/0.10}+0.20,

where vtv_t is the valid fraction within the scoring domain before imputation. Coverage and finite-value terms coincide here, combining their weights of 0.30 and 0.25; the constant combines the default metadata contribution of 0.10 and an offset of 0.10.

For SST, normalization clips (x−l)/(u−l)(x-l)/(u-l) to [0,1], using training-only 0.5/99.5 percentile bounds; SIC uses fixed [0,1] bounds. The innovation δt\delta_t averages absolute differences between consecutive normalized fields over jointly valid in-domain cells. It is zero at the first dataset date or when no valid pair exists. The scoring domain is the supplied mask, or the full grid if none is supplied. All quantities use historical observations; a window's first descriptor can use the preceding date.

At complete coverage, qt=0.75+0.25e−δt/0.10q_t=0.75+0.25e^{-\delta_t/0.10}. Its variation reflects temporal consistency, including genuine field evolution. A date-level scalar cannot locate regional corruption. Shared q scales all writes, while local features determine spatially varying gates. A small corrupted region may barely change q, making its treatment dependent on those features. A spatial confidence map would provide explicit regional quality information. This construction and its scope will be included in Appendix B.3.

W4. Isolating Field Memory under matched information and supervision.

A4. Field Memory couples retention, writing, and propagation under phase, reliability, and interface conditioning. The no-phase ablation zeros the five calendar features and retrains with other components and losses unchanged. Its increase from 0.9390 to 1.5392°C measures phase within this design; comparison with another architecture cannot identify that architecture's benefit from phase.

We therefore project and broadcast the same five phase features into each baseline, keeping its backbone, losses, and training protocol fixed.

Table P. Effect of explicit phase. Area RMSE in 10−2∘C10^{-2}{}^\circ\mathrm{C}.

ModelWithout explicit phaseWith explicit phase
3D U-Net102.0101.8
Earthformer103.4103.1
STDMamba117.0116.5
FORMA153.993.9

These controls support the value of explicit phase within FORMA's coupled architecture. We understand why its unusually large gain attracts scrutiny; our extensive experiments have not yet fully explained its magnitude or when comparable gains arise. We report the observation and controls faithfully.

The original RWKV control changes conditioning and propagation together. To evaluate the complete mechanism with information held fixed, we replace the full Field Memory block with conditioned ConvLSTM. Both share field/validity, phase, q, and interface features through the same input-adaptation design, encoder/decoder, interface supervision, all five losses and weights, splits, and checkpoint selection. Each trains its own parameters.

The temporal-only control retains the encoder, temporal accumulators, input mixing, retention and write gates, and decoder. It removes only the memory-transport term while keeping the retention coefficient fixed.

Table R1. Matched information and supervision. Errors: 10−2∘C10^{-2}{}^\circ\mathrm{C}.

Memory mechanismArea RMSEInterface RMSEBasin error
Conditioned temporal-only108.7127.46.2
Conditioned ConvLSTM104.9121.65.9
FORMA with self routes106.8124.16.1
Full Field Memory93.9109.65.5

The 10.5% gain over ConvLSTM evaluates the complete mechanism. Self routes replace global/local operators by their row-sum diagonals while keeping retention, writing, input mixing, supervision, and route coefficients fixed; their 12.1% gap examines spatial exchange within it. These comparisons will be included in the revised ablations.

AuthorsOpenReview ↗

Response to Reviewer 7GHq (Part 3)

W5. Training variability and the scope of repeated experiments.

A5. We chose to devote our training budget to comparisons across architectures, two geophysical fields, and mechanism ablations. Repeating this suite entails substantial additional training. We therefore favor targeted repetitions that test the central claims; we would not regard an automatic expansion to every experimental setting as proportionate to this study.

Your concern is particularly relevant to SIC IIEE: 104.2 versus IceMamba's 104.6 is a 0.38% difference in individual runs. The concentration-RMSE and SIE-error reductions are 4.3% and 25.4%. We report the IIEE ordering as an observed numerical difference, without assigning statistical significance.

To examine this at a bounded cost, we use a focused three-seed comparison with fixed splits and validation-based checkpoint selection. SD is the sample standard deviation across the three runs. Units follow Table 3.

ModelSubmitted RMSE / SIE / IIEEThree-seed mean ± sample SD (RMSE / SIE / IIEE)
IceMamba11.7 / 39.0 / 104.611.74 ± .11 / 39.20 ± .80 / 104.70 ± .82
FORMA11.2 / 29.1 / 104.211.22 ± .09 / 29.30 ± .56 / 104.40 ± .36

The mean reductions are 4.4% for concentration RMSE and 25.3% for SIE error. IIEE differs by 0.30, with the ordering reversing in one of the three paired runs. These results support emphasizing concentration and extent improvements while treating the small IIEE difference cautiously. Three repetitions characterize a limited sample of training variability; they do not establish significance.

Q1. How does FORMA distinguish genuine interface movement from observation noise, and what controlled evidence supports this distinction?

A1. The paired diagnostic in W1/A1 directly addresses your example: coherent motion and incoherent corruption have the same phase, coverage, and mean absolute innovation, hence the same q=.9176. The model therefore cannot distinguish them by lowering q for one case. Its response must use differences in the observed spatial features and the resulting recurrent history.

Table D1 illustrates that distinction. Mean Γ is .386 for coherent motion and .164 for corruption, while D is .742 and .891, respectively. Both write values obey Γ≤q. These gates are not probabilities of physical motion: D scales the retained state, while Γ scales newly accumulated evidence. Higher Γ alone would not establish a better forecast, because the evidence amplitude and spatial transport also matter.

We therefore include a uniform-write intervention on the same frozen model. It preserves each channel's mean write coefficient within the predefined band but removes its spatial allocation. The RMSE changes from .046 to .058 for motion and from .061 to .077 for corruption. Each comparison uses that case's own clean target; the difference between the two cases' raw RMSE is not a measure of discrimination. This control tests the usefulness of spatially selective writing, while matching q tests whether the response goes beyond a global reliability scale. The two cases hold phase fixed, so this diagnostic does not quantify every interaction among all three conditioning signals.

Q2. What are the component scores and normalization of q?

A2. W3/A3 gives the merged formula. For reproducibility, the reference setting can be written component by component as

\quad s_t=e^{-\delta_t/0.10}.$$ Here coverage c and finite-value validity f coincide with the pre-imputation valid fraction v; the metadata component takes its default value m=1. Thus their contributions reduce to 0.55v+0.25s+0.20. First retain the validity mask, then normalize the observed field using training-only SST percentile bounds or fixed SIC bounds. Compute δ as the mean absolute difference on jointly valid cells at consecutive dates. This averages over the supplied scoring mask, or the full grid when no mask is supplied. At the first dataset date or without valid pairs, δ=0. Finally evaluate the exponential and weighted sum, and broadcast the scalar to spatial locations. These steps use only observations available at that date. For v=1 and δ=.04, s=.6703 and q=.9176. This is a deterministic calculation, not an empirical reliability measurement. The default δ=0 convention gives the consistency term its maximum; with no valid pairs that is a boundary convention, not evidence that the observation is trustworthy. This component-level specification will be included in Appendix B.3.
Reviewer_9rG2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FORMA for 90-day SST and SIC forecasting. It introduces Field Memory, a spatial state updated via an RWKV-style recurrence. The update is conditioned on seasonal phase, observation reliability, and interface features. Experiments on OISST and NSIDC SIC show lower errors than several baselines

Critical strengths

Separating seasonal context, data quality, and sharp fronts is well-motivated for geophysical fields.

Transferring the exact same architecture from SST to Arctic SIC demonstrates versatility.

The authors provide a mathematical bound on the state transition operator in Appendix C.

Critical weaknesses
  1. Unfair baseline setup : In Table 4(a), removing the calendar phase (No phase conditioning) causes FORMA's RMSE to collapse from 93.9 to 153.9. This is worse than almost all baselines in Table 1. Yet Table B.1 states that baselines only receive field/validity input or single-field OISST input, with no mention of the calendar phase . Moreover, FORMA is trained with a 5-term loss (Eq. 12) that directly supervises future inter faces and regional means. These directly match the evaluation metrics Interface RMSE and Basin error. Training baselines without or these auxiliary losses makes Table 1 invalid.
  2. The evaluation of seasonal prediction baselines and anomaly-based metrics is insufficient. The metric referred to as "ACC" in the appendix is ​​calculated by subtracting the spatial mean from the raw fields, rather than subtracting the location-specific seasonal climatology. As such, it measures the correlation of spatial patterns and does not directly reflect the model's ability to predict anomalies. I recommend including skill scores relative to seasonal climatology, anomaly correlations, and results broken down by season or region. 3.The paper consistently employs 20 epochs across all models; however, a fixed number of epochs does not guarantee that different models reach a comparable level of convergence.
  3. The rationale for selecting OISST as a proxy for observational reliability is unclear. OISST v2.1 is a "gap-free" Level-4 interpolated analysis product, rather than raw satellite swaths containing missing data.
  4. The novelty of the proposed method, relative to existing recurrent state models, requires further clarification. The potential contribution lies in the specific combination of feedback signals and two types of spatial propagation, rather than in the concept of field memory itself.
Questions for authors

How much of an advantage does FORMA retain when provided with the same seasonal and reliability information as comparable RWKV, ConvLSTM, or other recurrent baselines?

Relative to seasonal climatology and anomaly persistence, what is the model’s predictive skill at 30, 60, and 90 days, respectively?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9rG2

Dear Reviewer 9rG2,

We sincerely appreciate your time and effort in reviewing our paper. Thank you for your positive assessment of our research motivation, particularly recognizing that explicitly accounting for seasonal context, data quality, and sharp fronts is well motivated for geophysical forecasting. We are glad that you recognized the versatility demonstrated by transferring the same architecture from SST to Arctic SIC. We also appreciate your recognition of the mathematical bound on the state-transition operator in Appendix C. We hope our responses below address your concerns. We would greatly appreciate it if you would consider updating your score in light of these clarifications and results.

W1. Fairness of calendar inputs and auxiliary supervision.

A1. The no-phase ablation tests phase within FORMA's integrated design: we retrain with five calendar features set to zero, keeping other components and losses unchanged. RMSE rises from 0.9390 to 1.5392°C. Comparing this variant with other architectures does not isolate their benefit from calendar information.

To test that benefit, we project and broadcast the same five phase features into each baseline's inputs, keeping its backbone, losses, and training protocol unchanged.

Table P. Effect of explicit phase. Area RMSE in 10−2∘C10^{-2}{}^\circ\mathrm{C}.

ModelWithout explicit phaseWith explicit phase
3D U-Net102.0101.8
Earthformer103.4103.1
STDMamba117.0116.5
FORMA153.993.9

We understand your concern about the unusually large phase-ablation gap. Table P shows that explicit seasonal information benefits FORMA much more than the tested baselines. This suggests that it may be particularly effective in certain architectures, rather than a generally transferable improvement for geophysical forecasting.

We have conducted extensive experiments and consulted related studies to understand this difference. Zeng et al. [1, Table 6] find that removing timestamp embeddings affects Informer, Autoformer, and FEDformer quite differently. TimeXer [2, Section 4.2] shows that the way external variables are incorporated also affects their benefit. These findings offer possible explanations involving temporal inductive biases and conditioning design. However, our experiments and this literature have not yet established why FORMA gains so much from explicit phase, or when a similar gain should be expected elsewhere. We can therefore report the observation and comparisons faithfully, while acknowledging that its underlying explanation remains incomplete. This discussion will be included in the revised ablation section.

Table P retains each baseline’s original losses and training protocol to evaluate the effect of adding the same phase information. Table R1 matches the conditioning inputs, all five losses, and their weights to evaluate the contribution of the memory mechanism; the configuration is detailed in Q1.

We noticed two points numbered 3. We answer the metric and training questions in W2(a) and W2(b), followed by the OISST question in W3.

References

[1] Zeng et al. Are Transformers Effective for Time Series Forecasting? AAAI 2023. https://arxiv.org/abs/2205.13504

[2] Wang et al. TimeXer. NeurIPS 2024. https://arxiv.org/abs/2402.19072

AuthorsOpenReview ↗

Response to Reviewer 9rG2 (Continued)

W2(a). Interpretation of ACC and evaluation of anomaly skill.

A2(a). Appendix B.2 explicitly defines the reported correlation by subtracting each map's area-weighted spatial mean. It measures spatial-pattern agreement, and Dcorr=103(1−ACC)D_{\mathrm{corr}}=10^3(1-\mathrm{ACC}) expresses its correlation deficit. Your distinction concerning climatological anomalies is correct: those anomalies require a location-specific seasonal reference. Ocean forecasting work such as NeuralOM [3] uses this climatological interpretation of ACC.

We will call the existing metric “spatial-pattern correlation” and report climatology-referenced anomaly correlation separately. This preserves the original evaluation while making the two quantities unambiguous. The additional metric uses location-specific day-of-year means estimated exclusively from the 1982–2018 training period. Q2 reports the lead-day evaluation under this definition.

W2(b). Whether 20 epochs provides adequate convergence.

A2(b). Training budgets in geophysical forecasting reflect the target task. For example, Pangu-Weather [4] trains its multivariate atmospheric models for 100 epochs. Our experiments forecast SST or SIC separately, with slowly varying background fields alongside moving fronts and ice edges. For these configurations, all models attain their best validation area RMSE before epoch 20, with no material improvement over the final several epochs. These convergence observations support 20 epochs as an adequate budget for our experiments.

Each model is evaluated at its best validation checkpoint. This is consistent with the validation-based early stopping and best-checkpoint evaluation in the official SST forecasting implementation of STDMamba [5]. The convergence evidence and checkpoint-selection procedure will be included in the revised experimental details.

W3. Why use reliability conditioning with gap-free OISST?

A3. Thank you for raising this question. Our q is computed from historical field validity and temporal consistency to regulate the acceptance of incoming evidence.

OISST's gap-free coverage is obtained through interpolation; uncertainty still varies with the observations supporting the analysis. NOAA supplies an interpolation-error field as a confidence or quality measure [6], and v2.1 improves observation inputs and bias correction [7]. Complete coverage therefore does not remove the motivation for reliability-conditioned updates. Our q provides a field-derived conditioning signal, rather than an estimate of NOAA's interpolation error: qt=0.55vt+0.25e−δt/0.10+0.20,q_t=0.55v_t+0.25e^{-\delta_t/0.10}+0.20, where vtv_t is the valid fraction within the scoring domain before imputation and δt\delta_t is the mean absolute change between consecutive normalized fields over jointly valid cells. The two validity terms combine into 0.55vt0.55v_t; the remaining constant includes the default metadata contribution.

At fixed coverage, q varies with temporal consistency, which can reflect both real field evolution and input corruption. The clean-history ablation measures the predictive value of this conditioning, while the synthetic-gap and outlier tests examine responses to controlled corruption with q recomputed from the perturbed history. This explanation and construction will be included in revised Appendix B.3.

References

[3] Gao et al. NeuralOM. AAAI 2026. https://arxiv.org/abs/2505.21020

[4] Bi et al. Pangu-Weather. Nature, 2023. https://doi.org/10.1038/s41586-023-06185-3

[5] Jiang et al. STDMamba. IEEE TGRS, 2025. https://github.com/XDjiang25/STDMamba

[6] NOAA/NCEI. OISST, Additional Data. https://www.ncei.noaa.gov/products/optimum-interpolation-sst

[7] Huang et al. DOISST v2.1 improvements. J. Climate, 2021. https://doi.org/10.1175/JCLI-D-20-0166.1

AuthorsOpenReview ↗

Response to Reviewer 9rG2 (Final Part)

W4. Clarify the novelty of Field Memory.

A4. Thank you for this insightful comment. The central contribution is the coupling itself: Field Memory makes long-term retention, acceptance of new observations, and global/local spatial propagation jointly adaptive within one predictive state. This gives the model a coordinated way to preserve persistent seasonal structure while updating evolving interfaces. We call this integrated formulation Field Memory.

Table 4(a) supports its effectiveness. Removing phase, reliability, or the interface components raises area RMSE from 93.9 to 153.9, 120.7, or 112.8, respectively; static gates reach 157.2. The full design achieves the best accuracy among these variants, supporting the value of these components working together in the state update. Q1 further examines spatial propagation with matched information and supervision. Equation (11) and Appendix C analyze this same state, tracing past writes and bounding propagation under fixed conditioning. This explanation will be included in the revised method section.

Q1. What advantage remains with the same information and training objectives?

A1. We compare controls sharing field/validity inputs, phase, the same field-derived q, interface-feature construction, and all five losses with identical weights. The temporal-only control retains FORMA's encoder, temporal accumulators, input mixing, retention/write gates, and decoder. It removes only the memory-transport term, keeping the retention coefficient fixed. ConvLSTM receives the same conditioning features as recurrent inputs.

The self-route control replaces the global and local propagation operators with their row-sum diagonals, retaining the retention gate, writing, input mixing, supervision, and routing coefficients.

Table R1. Matched information and supervision. Errors in 10−2∘C10^{-2}{}^\circ\mathrm{C}.

ModelArea RMSEInterface RMSEBasin error
Conditioned temporal-only108.7127.46.2
Conditioned ConvLSTM104.9121.65.9
FORMA with self routes106.8124.16.1
Full Field Memory93.9109.65.5

The full model reduces area RMSE by 13.6%, 10.5%, and 12.1% relative to the temporal-only, conditioned ConvLSTM, and self-route controls, respectively. We will add these matched controls and configurations.

Q2. What is the predictive skill at days 30, 60, and 90?

A2. We subtract the training-only seasonal means from predictions and observations. Anomaly correlation is their area-weighted normalized inner product, averaged over test origins.

Table R2. Lead-day SST evaluation. RMSE in 10−2∘C10^{-2}{}^\circ\mathrm{C}. Table 1's 93.9 summarizes all 90 forecast days; these rows evaluate individual lead days.

Lead dayArea RMSEAnomaly correlation
3087.5.694
6099.1.572
90111.6.390

For SSb=1−MSEFORMA/MSEbSS_b=1-\mathrm{MSE}_{FORMA}/\mathrm{MSE}_b, the scores against seasonal climatology are .429/.275/.092 at days 30/60/90; against persistence of the latest anomaly, they are −.030/.081/.095. We will report both comparisons alongside anomaly correlation using the same dates, masks, and weights.

Thank you for your careful review and constructive suggestions. We will further detail the matched experimental settings, anomaly-based evaluation, and reliability-score computation in the revised manuscript and supplementary materials to improve clarity and reproducibility. All comparisons use the same test dates, masks, and area weights.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Response to Common Concerns

Summary of Revisions and Response to Common Concerns

We sincerely thank all reviewers for their careful reading and continued discussion. We appreciate 7GHq’s recognition of preserving temporal information while adapting to evolving boundaries and observation quality, and complementary global/local interactions with linear complexity at fixed width and rank. We thank 9rG2 for recognizing the motivation for seasonal, quality, and interface conditioning and the versatility of the same architecture across SST and Arctic SIC. We also appreciate HdHV’s acknowledgment of the clean conceptual design of retention, writing, and local routing, and consistent improvements in field accuracy and front/ice-edge localization.

We have uploaded a revised manuscript, with textual additions and changes highlighted in blue.

1. Changes to the paper

  • Information flow (7GHq; Appendix E.1): Clarified temporal accumulators, evidence readout, spatial memory, and location-wise gates; distinguished input mixing from memory propagation.
  • Reliability conditioning (all reviewers; Sec. 3.2, Appendix E.2): Specified q’s components, normalization, masks, defaults, and weights, and explained its interaction with spatially varying gates and its interpretation on gap-free analyses.
  • Relation to prior methods (9rG2, HdHV; Appendix E.3): Expanded comparisons with TrajGRU and Efficient Attention, explaining how current observations regulate evidence acceptance and redistribution of accumulated memory.
  • Propagation analysis (Appendix E.4): Expanded the derivation for normalized routes and continuing writes; distinguished conditional state contraction from sensitivity to changed observations.
  • Experimental interpretation (all reviewers; Secs. 4.1, 4.4, Appendix E.5): Clarified spatial-pattern versus anomaly correlation, the scope of grouped ablations, validation-based checkpoint selection, and independent training of the same architecture on SST and SIC.

2. Contribution and evidence from the discussion

Architectural contribution. Field Memory couples temporal evidence accumulation with feedback-conditioned spatial propagation. Temporal accumulators produce evidence at each location; a separate spatial state retains and exchanges earlier writes through global and local routes. Subsequent observations can therefore change how accumulated evidence propagates. Appendix E.3 positions this formulation relative to adaptive recurrence and factorized attention.

Matched comparisons. Table R1 in our responses matches conditioning inputs and all five losses and weights. FORMA achieves area RMSE 93.9 versus 104.9 for conditioned ConvLSTM, 108.7 for the temporal-only control, and 106.8 for self routes: reductions of 10.5%, 13.6%, and 12.1%, respectively. Self routes retain interface supervision and routing coefficients while removing spatial exchange. These controls evaluate the complete mechanism and spatial exchange under matched conditions. The conditioned ConvLSTM comparison shows that FORMA’s complete memory mechanism provides an additional gain under matched inputs and supervision. The self-route comparison more specifically supports the contribution of spatial exchange within that mechanism.

Phase conditioning and anomaly skill. Table P shows architecture-dependent benefits from explicit phase, supporting its effectiveness within FORMA’s coupled design. The unusually large gain still warrants further mechanistic study. Lead-day evaluations show positive skill against seasonal climatology at 30, 60, and 90 days, and against anomaly persistence at 60 and 90 days; anomaly persistence performs slightly better at day 30.

We thank 9rG2 for suggesting coupled multivariable inputs and probabilistic or ensemble forecasts; coupled-field and probabilistic forecasting are future directions in our conclusion. We also thank HdHV for the follow-up discussion on the architectural contribution and welcome 7GHq’s further feedback. We will continue to address subsequent comments and incorporate relevant clarifications into our work. Finally, we will release the source code to support reproducibility and enable broader application of FORMA in geophysical forecasting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.