acceptodds
Desk rejected at ICLR 2027

PhyWarp: Parameter-Efficient Adaptation and Knowledge Transfer for Tropical Cyclone Forecasting

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series foundation models (TSFMs) offer reusable forecasting priors that can be adapted to downstream tasks without training dedicated models from scratch. However, tropical cyclone (TC) forecasting presents a challenging transfer setting: available storm records are limited, intensity and motion follow different dynamics, and storm evolution depends on multivariate environmental factors that vary across ocean basins. These characteristics limit the effectiveness of direct TSFM transfer. Moreover, the computational cost of TSFMs can hinder their deployment on resource-constrained meteorological devices. In this work, we introduce PhyWarp, a parameter-efficient framework that adapts frozen TSFMs to TC forecasting by representing endogenous storm states through multiple physically meaningful views and using exogenous environmental context to guide their combination. A shared frozen backbone produces forecasts from these complementary representations, which are combined through lightweight adaptation across lead times and prediction targets. To reduce computational cost, we further distill the adapted knowledge into a lightweight predictor. Experiments on Atlantic and eastern North Pacific data show consistent improvements across three representative TSFM backbones and lead times from 6 to 48 hours, while outperforming supervised time-series models. Ablation studies further demonstrate the complementary benefits of physically informed state representations and adaptive forecast fusion. The lightweight predictor retains strong forecasting performance while substantially reducing inference cost. The code and pretrained weights will be made publicly available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Tropical cyclone forecasting, Time-series foundation models, Parameter-efficient adaptation

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

The paper introduces PHYWARP, a framework that adapts frozen time-series foundation models for joint tropical cyclone track and intensity forecasting. PHYWARP's contributions include three complementary storm-state representations, environment-aware adaptive forecast fusion, an intensity correction module, and distillation into a compact predictor that does not require foundation-model inference.

The framework was validated using 5,819 SHIPS cases from the Atlantic and eastern North Pacific, as well as chronological train/validation/test splits, three foundation-model backbones, and six supervised baselines with 6–48-hour forecast horizons. The reported results show consistent improvements over direct frozen-model transfer; the best variants outperform the evaluated supervised baselines. Ablation studies support the distinct roles of the components, and distillation achieves near-teacher accuracy with approximately 93× faster CPU inference. While these findings establish the framework's effectiveness within the tested setting, its operational competitiveness and generalization to other basins remain unverified.

Critical strengths
  1. Clear Writing and Organization: The paper is clearly written and logically organized. The progression from storm-state representations to environment-aware fusion, residual correction, and distillation makes the proposed framework easy to follow.
  2. Relevant Problem Setting: Adapting general-purpose forecasting priors to tropical-cyclone storm-state prediction is a relevant research question. The distinction between motion, intensity, and environmental context provides an intuitive basis for exploring task-specific adaptation.
  3. Useful Diagnostic Analyses: The component ablations and horizon-dependent fusion-weight visualizations provide useful insights into the framework’s behavior. The inclusion of both scratch-trained and direct-distillation student controls also makes the distillation analysis informative.
  4. Explicit Separation of Training and Deployment: The TSFM-free student separates offline teacher adaptation from lightweight inference. Table 4 demonstrates substantial inference-time compression relative to the adapted teacher while retaining competitive forecasting accuracy.
Critical weaknesses

1. Disentangling TSFM Prior Gains from Task-Specific Adaptation

The scratch-trained GRU in Table 4 is already competitive with the adapted Chronos-2 teacher, particularly in short-range intensity and 48-hour track forecasting. Although PHYWARP-KD provides clear short-range track improvements over scratch, the existing controls do not fully establish how much of the advantage depends on pretrained forecasting priors rather than task-specific representations and supervised correction.

2. Establishing Task-Relevant Forecast Skill and Applicability

The evaluation includes general time-series models but lacks inexpensive meteorological references, such as intensity persistence or recent-motion extrapolation. Moreover, requiring complete 48-hour histories and targets through the following 48 hours—even for the 6-hour experiment—restricts the evaluated initialization opportunities. Pooled and basin-wise aggregate errors do not establish whether the improvements extend to important situations such as rapid intensification or strong storms.

Questions for authors

Q1. Student Capacity and the Intended Deployment Budget

Appendix A.3.5 specifies the student architecture, but could you explicitly report its parameter count for each forecast horizon and clarify the intended deployment budget, particularly memory and latency constraints? Since all student variants already run at the same reported CPU latency, the relevant comparison is between PHYWARP-KD and well-tuned, potentially modestly larger scratch-trained models—not only between the student and the TSFM teacher. Have you examined this accuracy–cost trade-off through a small capacity sweep under matched inputs and objectives? Please also distinguish deployment savings from the offline cost of teacher adaptation and distillation.

Q2. Event-Specific Verification and a Storm-Level Case Study

Could you provide targeted verification for at least one clearly defined high-impact subset, such as rapid intensification or strong storms, including the event definition, numbers of storms and forecast cases, and appropriate event-detection or conditional-error metrics? A held-out storm case comparing the observed evolution with the teacher, distilled student, scratch model, and a simple meteorological reference would also help reveal where the framework succeeds or fails. This would complement the aggregate forecast–observation scatterplot currently provided in Figure 4.

Reviewer_24cdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PHYWARP, a parameter-efficient framework for adapting frozen time-series foundation models to tropical-cyclone track and intensity forecasting. By combining physically informed representations with adaptive fusion of model predictions, PHYWARP leverages environmental conditions while preserving the pretrained backbone. The adapted model is then distilled into a lightweight GRU for efficient deployment without requiring TSFM inference. Experiments on Atlantic and eastern North Pacific SHIPS data show substantial improvements in track forecasting accuracy, reducing 6-h track MAE from 52.53 km to 20.45 km and improving pressure and wind predictions over six supervised baselines across all evaluated lead times.

Critical strengths
  1. Rather than modifying model weights, the paper adapts the input representation to the forecasting task. Since the frozen channel-independent backbone processes each variable independently, PHYWARP transforms storm states into three deterministic, invertible views that encode absolute conditions, temporal changes, and heading-relative motion. These representations recast cross-variable dependencies into patterns that the backbone can model more effectively.
  2. The fusion gate evaluates each candidate based on its forecast, deviation from the recent state, lead time, view identity, and environmental context, allowing view weights to reflect prediction-specific evidence rather than input features alone. This provides a more principled mixture-of-experts design.
Critical weaknesses
  1. The cyclone-specific contribution needs clearer validation. Most components can be applied to generic multivariate forecasting, while the domain-specific contribution mainly lies in the heading-based rotation and variable grouping. Since no ablation isolates the effect of heading rotation, the benefit of the proposed physical representation remains unclear. Comparing against an unrotated east–north representation and evaluating on another weather dataset would help establish its value.

  2. The methodological novelty could be better established. The framework combines established techniques, including multi-view representations, softmax gating, residual correction, and response-based distillation. Physical knowledge enters primarily through input reparameterization, without explicit physical constraints or dynamics modeling. Further analysis of when adaptive fusion outperforms simple averaging would help clarify the method's contribution beyond component integration.

  3. Figure 2 could communicate the architecture more clearly. The current figure omits tensor shapes, the temporal axis, and an explicit distinction between frozen and trainable parameters. It also does not clearly show the horizon-dependent residual correction, target-specific fusion weights, or training losses. Adding these details would make the architecture and its key design choices easier to understand.

  4. The baseline comparison could be more comprehensive. The evaluation lacks classical forecasting baselines such as persistence, seasonal naïve, ARIMA, and regression with exogenous covariates, making it difficult to assess the gains over simple references. Adding persistence and ARIMAX baselines would strengthen the comparison, while testing a natively multivariate backbone such as the recent TimesFM 3 would further examine the channel-independence assumption.

Questions for authors

See Weaknesses

Reviewer_LWekOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PhyWarp, a parameter-efficient framework for adapting frozen time-series foundation models (TSFMs) to tropical cyclone track and intensity forecasting. It combines three storm-state representations through environment-aware fusion and distills the adapted model into a lightweight student. Experiments on two ocean basins show improvements over frozen TSFMs and the evaluated supervised time-series baselines.

Critical strengths

The framework is clearly structured and evaluated across multiple backbones, forecast horizons, and ocean basins. The distillation results also demonstrate substantial inference speedup.

Critical weaknesses

Insufficient baseline comparisons. The experiments focus on general-purpose time-series models and TSFMs, without comparing against existing TSFM adaptation methods discussed in the paper or specialized AI and traditional TC forecasting methods. This limits the evidence for the method’s competitiveness and novelty.

Insufficiently substantiated motivation. The proposed multi-view representations and environment-aware fusion address relatively general challenges, but the paper does not clearly establish their necessity or distinguish their contributions from existing adaptation and exogenous-variable modeling approaches. The practical need for the substantial inference speedup is also insufficiently justified.

Questions for authors

Why are existing TSFM adaptation methods discussed in Section 2.3 not included as baselines?

What evidence demonstrates that the proposed design addresses specific limitations of direct TSFM transfer beyond improving forecasting performance?

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

Related papers

Discussion (0)

Sign in to comment.