TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness

Thank you for acknowledging the contributions of our ARP mechanism. We also understand your concerns regarding the motivation behind using both NWP data and the diffusion model. Below, we provide detailed responses to these questions, and we hope they will help alleviate some of your concerns about our method.

Q1: A strong motivation for using a diffusion model over other deep learning approaches (e.g., VAEs, Transformers) for precipitation forecasting is not clearly provided.

We would like to clarify the motivation for choosing diffusion models over other representation learning-based approaches.

1. TC precipitation exhibits chaotic behavior and inherent uncertainty, which makes probabilistic forecasting more suitable than deterministic approaches. This is supported by the theoretical discussion in Atmospheric Modeling, Data Assimilation and Predictability, particularly Section 6.5, which states:"The chaotic behavior of the atmosphere requires the replacement of single ‘deterministic’ forecasts by ‘ensembles’ of forecasts." Deterministic models like VAEs can be seen as single forecasts, whereas diffusion models naturally resemble ensemble forecasting by sampling multiple possible futures. This makes them well-aligned with the demands of real-world atmospheric prediction, especially under extreme conditions such as TCs.

2. Previous methods—including deterministic models like U-Net, VAEs, and Transformer-based architectures—often produce over-smoothed forecasts with limited fine-grained spatial details. We briefly discuss these issues in Lines 69–86 (left column) of the paper. In contrast, diffusion models naturally incorporate noise and denoising processes, allowing them to generate more detailed and realistic outputs.

Q2: The paper describes the use of physical processes only from the results of NWP, but incorporating more TC-related information as model input is an intuitive approach rather than one inspired by physical principles. (For Weaknesses 1 and 4)

It is true that our approach adopts a relatively simple yet effective strategy to integrate deep learning with NWP forecasts. However, to the best of our knowledge, this is the first work that combines deep learning with NWP data specifically for TC precipitation forecasting. While the idea may appear intuitive in hindsight, however, the novelty must be evaluated before the idea existed. The inventive novelty was to have the idea in the first place. If it is easy to explain and obvious in hindsight, this in no way diminishes the creativity (and novelty) of the idea.

In our approach, we do not directly embed physical equations into the model (such as PINNs), but instead let the model learn the embedded physical knowledge present in the IFS forecast. This is a simple yet practical alternative to explicitly encoding physical constraints, especially considering the complexity of modeling TC precipitation—a highly nonlinear and chaotic process.

That said, we fully agree that incorporating meteorological physics in a more explicit way (e.g., through constraint-aware loss functions or hybrid physics-ML models) is a promising direction. Bridging the gap between data-driven modeling and physical interpretability is part of our long-term research vision.

Q3: Is incorporating NWP outputs in the model truly reasonable?

Reviewer Ndz5 raised a similar concern, which we responded to in Q1. For a more detailed explanation, please refer to that response.

Q4: The ablation study of individual encoders (historical vs. future data) is missing. Additionally, while different lead times are tested, a deeper analysis of how performance degrades over time is missing.

We would like to clarify that ablation experiments on the historical and future data encoders were indeed included in Table 3. Specifically, the third row in Table 3 represents our model with the ARP and multimodal encoder (M), but without the historical encoders and future data encoders. The fourth row shows the full version of our model, which includes both historical and future encoders. The performance improvement between these two rows demonstrates the effectiveness of integrating NWP forecasts and our encoder designs.

Regarding performance degradation over time, it is a well-known challenge in sequential forecasting tasks. In our setting, this issue is particularly pronounced due to the chaotic nature of atmospheric systems, where small disturbances can amplify over time and lead to substantial divergence in predictions (as described by the butterfly effect). We will provide a more comprehensive analysis in Section D.1 (Line 755) of the camera-ready version.

Q5: The performances of our method on different TC lifecycle stages and intensities.

The results are shown at https://limewire.com/d/HHv6Z#HXP582VLSv. Please refer to this link for detailed results.

If you have more questions, We'd like to discuss them with you during the author-reviewer discussion period.

Thank you for recognizing the value of integrating our method with NWP. We also understand your concerns regarding the potential impact of using NWP data on the flexibility of our approach. These insights will serve as valuable guidance for our future research. Below, we provide responses to your questions, and we hope they will help address some of your concerns about our method.

Q1: The flexibility of the method is limited to some extent due to the use of NWP data.

We appreciate the reviewer’s concern regarding the flexibility and real-world applicability of our method.

1. Please note that our model also provides a version that does not rely on IFS forecasts. As shown in Table 3 (Row 3), our model without IFS still outperforms other deep learning baselines listed in Table 1. This is also discussed in Lines 413-418 (right column) of the manuscript. Therefore, our approach remains functional and competitive even in the absence of IFS inputs.
2. ECMWF has announced that it will make its forecast data fully open starting in October 2025 (https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwf-achieve-fully-open-data-status-2025). This will significantly reduce the difficulty of accessing IFS forecasts, making it increasingly feasible to incorporate them in practice. We believe our work offers a timely exploration of how to leverage such forecasts effectively in deep learning frameworks.
3. we emphasize that IFS and deep learning are not mutually exclusive; rather, their integration can lead to more accurate and efficient forecasts. To our knowledge, this work is the first to integrate NWP forecasts with deep learning for TC precipitation prediction. While IFS captures physical laws through numerical simulation, deep learning excels at learning hidden patterns from data. Our design—which directly feeds IFS outputs into a dedicated encoder for representation learning—offers a simple yet effective fusion strategy. This approach may serve as a useful reference for future efforts in combining NWP with data-driven models across various forecasting tasks.

Q2: The difference with DiffCast, Pangu and Fengwu.

We would like to clarify the unique contributions of our model and how they differ from prior works:

1. Difference from DiffCast: While both our approach and DiffCast involve residual prediction, the underlying concepts are fundamentally different. In our model, we view precipitation forecasting as an accumulative process of rainfall changes over time, where the model directly learns to generate the adjacent residual (future rainfall change) rather than the absolute future value. This is reflected in lines 102-108 (left column).

   In contrast, DiffCast models the residual between a deterministic forecast and the ground truth, as shown in Section 4.2 of their paper. Their diffusion module acts more as a correction mechanism to add details to the output of a deterministic backbone, rather than directly modeling rainfall evolution through temporal differences. Although DiffCast is an inspiring work, our method tackles a different formulation and learning objective. We will discuss this paper in the camera-ready version.

2. Difference from Pangu and Fengwu: The motivation for using multimodal meteorological inputs in our work is distinct. The prediction targets of Pangu and Fengwu are the future states of multiple variables themselves, hence multimodal inputs are a natural part of the task.

   In contrast, we focus solely on predicting TC rainfall, and we incorporate multimodal inputs to compensate for the limitations of using rainfall data alone, especially in capturing TC rainfall dynamics. Our design emphasizes the integration of environmental and physical information to enhance prediction quality, addressing a gap in existing regular precipitation forecasting models.

3. Empirical Validation: As shown in our ablation results (Table 3), both the proposed ARP and the multimodal input design (M) contribute significantly to model performance. This supports the value of our innovations, particularly in the context of TC precipitation forecasting.

Q3: It is necessary to visualize the residual prediction and show the contribution of how it can reduce cumulative errors.

The residual prediction visualizations are shown at this link: https://limewire.com/d/HHv6Z#HXP582VLSv. Please refer to this link for detailed results.

Regarding its role in reducing cumulative errors, predicting residuals rather than absolute rainfall values allows the model to estimate smaller relative shifts at each time step. As a result, even if there are minor prediction errors at individual steps, their cumulative impact is less severe. This approach helps mitigate long-term drift and leads to more stable multi-step forecasts.

Owing to space constraints, we would be glad to elaborate further on Methods and Evaluation Criteria (Point 2) or other questions during the author-reviewer discussion period.

Thank you for recognizing our work and for your valuable comments and stylistic recommendations. These suggestions will help us further improve the quality of this manuscript. We will incorporate the corresponding revisions in the camera-ready version. Below are our responses to the issues you raised.

Q1:Would it be possible to provide access to at least a subset of the dataset or a simplified version of the model during the review process?

A1:We have created an anonymous GitHub repository (https://anonymous.4open.science/r/TCP-Diffusion-ICML-review/README.md) that includes the testing code and a subset of the test data for our method.

Q2:Have the authors considered benchmarking their model against probabilistic forecasting techniques, especially in terms of uncertainty quantification?

A2:We have compared our method with PreDiff, which is also a probabilistic forecasting approach. While we have not explicitly benchmarked uncertainty quantification metrics against PreDiff at this stage, we have evaluated the stability of our predictions. As shown in Table 5, the standard deviation of our method's results is notably small, indicating that our model performs consistently across the dataset.

Q3:How adaptable is TCP-Diffusion to future improvements in NWP methodologies? Would retraining be necessary with every update to the NWP model, or can the framework accommodate updated inputs dynamically?

A3:Currently, our model does not dynamically adapt to different NWP methodologies. This is primarily because different NWP models may produce outputs with varying resolutions, spatial-temporal scales, and embedded physical assumptions. Existing offline deep learning frameworks, including ours, typically require consistent input characteristics and cannot yet generalize across heterogeneous NWP outputs without retraining. We greatly appreciate this insightful suggestion—it highlights a direction for making our work more practically applicable. In future research, we plan to incorporate adaptability to diverse NWP models into our framework design.

Q4:Could the authors clarify whether the model also improves fine-scale precipitation patterns, or does it tend to smooth out small-scale variability?

A4:Compared to deterministic forecasting methods, our model provides richer details in the prediction of fine-scale precipitation patterns. As illustrated in Figures 3 and 9, especially around the TC center, our method is capable of capturing the shape and structure of heavy rainfall regions. In contrast, deterministic baselines such as U-Net tend to smooth out these localized high-intensity features. This demonstrates the advantage of our probabilistic diffusion-based approach in preserving small-scale variability.

Q5:Could the authors provide insights into whether there is a feasible way to optimize the model's computational efficiency while maintaining its accuracy?

A5:As shown in Table 4, our method requires longer training time compared to some traditional deep learning models, and the inference time is also slightly higher than that of non-diffusion approaches. However, given the complexity of the TC precipitation forecasting task, we believe the computational cost of our model remains reasonable. Furthermore, our method is significantly more computationally efficient than NWP systems.

In terms of model design, we adopt a hybrid architecture combining CNNs and Transformers. For high-dimensional data, we utilize CNNs, which are computationally efficient and lightweight. For lower-dimensional inputs, we use Transformers, which, though more computationally intensive, offer stronger feature extraction capabilities—allowing for a more fine-grained understanding.

Additionally, we are exploring model transfer strategies, where a pre-trained large model is adapted to downstream tasks like TC precipitation forecasting. This would allow us to fine-tune only a small portion of the parameters while leveraging the representational power of the large model. This direction represents a promising way to further enhance computational efficiency and is a focus of our future research.

Q6: Does the model show any systematic underprediction or overprediction of extreme precipitation events?

A6: We conducted a focused evaluation on extreme precipitation events, which account for approximately 6.73% of the total test samples. The results show that our model tends to slightly overpredict these events, with an average overestimation of about 0.206 mm per 3 hours.

These findings provide valuable insights for future improvements. In particular, we plan to explore techniques such as weighted loss functions or targeted fine-tuning to better handle rare but high-impact precipitation extremes. Addressing this challenge is critical for enhancing the model’s reliability under severe weather conditions.

If you have more questions, We'd like to discuss them with you during the author-reviewer discussion period.

Thank you for recognizing our work, including the novelty of the task itself, the comprehensiveness of our experiments, and the clarity of the manuscript. We also understand the reviewer’s concerns regarding our use of IFS data. Below are our responses to some of the issues raised, and we hope they can help alleviate some of your concerns.

Q1: The model relies a lot on the IFS forecast from ECMWF.

1. Comparing the first row in Table 1 with the third row in Table 3 (the model version without IFS), we can observe that our method still slightly outperforms the persistence baseline overall. Furthermore, even without IFS inputs, our model achieves better performance than other deep learning-based baselines, as also mentioned in Lines 413-418 (right column) of the manuscript.

2. Our work is, to our knowledge, the first to integrate deep learning methods with NWP (specifically IFS data) for TC precipitation forecasting. As one of the contributions of this work, integrating with NWP can improve the performance of our model, which supports the effectiveness of this contribution and offers useful insights for broader weather forecasting tasks.

3. The acquisition and usage of IFS data will be further simplified. Please refer to the details in the response to the Q1 of Reviewer Ndz5, point 2.

Q2: The ablation study of ARP is currently missing.

It is possible that our explanation of Table 3 was not sufficiently clear. In fact, the ablation study of the ARP mechanism is already presented in the first and second rows of Table 3. The first row corresponds to the original baseline of our method without ARP, while the second row shows the results after incorporating the ARP mechanism. The performance improvement demonstrates the effectiveness of ARP.

Q3: Using ARP is a well-known technique for weather forecast see i.e., GenCast

We appreciate the reviewer for bringing GenCast (published in December 2024) to our attention. We have carefully studied the paper and will cite and briefly discuss it in our camera-ready version. Notably, we found that the Residual Prediction in GenCast shares conceptual similarities with our ARP module, which further supports the value of residual modeling in weather-related prediction tasks.

Our ARP design was inspired by the denoising process in diffusion models. Just as diffusion models iteratively refine predictions by removing noise, we view precipitation forecasting as a residual step-by-step accumulation process, as discussed in Lines 102-108 (left column) of our paper. We later found theoretical support for this idea in Chapter 6 of Atmospheric Modeling, Data Assimilation and Predictability, which further reinforced the motivation for adopting ARP in our framework. To the best of our knowledge, this is the first work to apply the ARP mechanism to the task of TC precipitation forecasting. We believe our work can provide inspiration for future research on specialized weather forecasting tasks.

Q4: Compare with global forecast model FuXi.(For the 2nd point of Methods And Evaluation Criteria)

We appreciate the reviewer’s suggestion and have supplemented additional experiments to address this point. Among existing large global forecasting models, FuXi (FuXi: a cascade machine learning forecasting system for 15-day global weather forecast) is currently the only one capable of providing global-scale precipitation predictions. However, FuXi does not publicly release official predictions for the total precipitation (TP) variable. So, we reproduced the FuXi model using the official codebase and evaluated its performance on TC precipitation. From the results, we observe that FuXi performs poorly and tends to significantly overestimate rainfall intensity. More details are shown at the link:https://limewire.com/d/HHv6Z#HXP582VLSv

Q5: Concern about the practicality of the approach for real-world scenario. (For the 1st point of Methods And Evaluation Criteria)

Regarding ERA5, although it is a reanalysis product, near real-time access is possible through collaboration with ECMWF. For instance, Pangu-Weather also utilizes ERA5 reanalysis data and has already been adopted within ECMWF's operational forecasting system for real-time prediction. Therefore, we believe the use of ERA5 in research and applied forecasting scenarios is reasonable and increasingly practical.

For IFS data, the generation time is shorter than that of ERA5 data. Moreover, since we use low-resolution IFS data, the required time is further reduced. This has been mentioned in Lines 427-428 (left column).

Q6: How ARP is going to reduce accumulative errors?

A similar concern was raised by Reviewer Ndz5, and we have addressed it in our response to their Q3. Please refer to that response for detailed clarification.

Owing to space constraints, we would be glad to elaborate further on Experimental Designs Or Analyses (Point 2) or other questions during the author-reviewer discussion period.