Multi-modal graph neural networks for localized off-grid weather forecasting

Firstly, I believe the paper's quality has not yet reached the standard of a top conference like ICLR. I suggest the author consider submitting to a more suitable venue or work towards a more impactful contribution to the topic, rather than an incremental improvement. I have shown some weakness that I can find in the paper writting aspect but the essential problem of this paper is that lacks breakthrough novelty.

1. The Introduction section lacks a discussion of the motivation. It does not systematically explain the motivation behind the work, nor does it thoroughly compare the proposed model with previous models.
2. The summary of contributions at the end of the Introduction has a high degree of repetition with earlier sections and is not concise or well-summarized.
3. The innovative aspect of the model is restrictive. The paper only uses local observation data to adjust global forecasts, which is restrictive.
4. The presentation of Figure 1 is not closely connected to the context and does not provoke further thought. It is merely shown without being integrated into the discussion. The symbol representing time in Figure 2 is also a bit small.
5. The related work section only briefly lists some relevant works without providing a systematic or organized analysis. Specifically, in the Gridded Weather Forecasting section, only two types of machine learning methods are listed without explaining their advantages or improvements. It also fails to highlight how the proposed model builds on or innovates beyond previous models.
6. The Discussion and Conclusion sections are not concise enough. Some paragraphs are overly lengthy, especially in the experimental results section where certain analyses are repeated, making the content appear redundant.
7. The paper uses many technical terms related to weather forecasting, which may be difficult for cross-disciplinary readers to understand due to a lack of background information.
8. The paper does not mention the conditions for actual deployment or the hardware resources required, making it hard to replicate the results.
9. The paper does not discuss limitations, nor does it analyze the computational costs, which affects the credibility of the work.
10. The Method section lacks an analysis of the model architecture, making it difficult for readers to understand the internal structure of the model.
11. The paper places the model architecture diagram in the Appendix, which makes it inconvenient for readers and easy to overlook. The author should reorganize the content of the paper.
12. The paper only describes the meaning of each term in the equations (eqs. 6 and 10) without explaining the overall logic of the equations. There is no explanation of the internal relationships between the equations. The explanations of equations in the Method section are isolated and not interconnected.
13. The paper's baseline is incomplete and does not include state-of-the-art approaches.

1. As weather station forecasting is essentially a spatio-temporal prediction task, the introduction of GNNs and heterogeneous graph for this type of problem is not novel that has been extensively explored in prior research [1].
2. The validation is very weak that only a few simple interpolation and persistence methods are compared. Incorporating more powerful numerical weather prediction and machine learning methods would strengthen the evaluation.
3. The absence of physical constraints in the proposed model raises significant concerns about the reliability and robustness of the model.
4. No code and datasets are provided, making this paper difficult to replicate.

[1] Spatio-temporal graph neural networks for predictive learning in urban computing: A survey. TKDE, 2023.

1. What innovative aspects does this article's model possess? Is it solely the application of heterogeneous graph networks to weather forecasting?

2. What is the distinction between the proposed dataset and existing datasets is a key contribution of this work？Given that the dataset is the primary innovation, the main text should include a more comprehensive introduction and analysis of its unique characteristics.

3. The paper lacks experiments ( such as experiments on current GNN baselines ), and the volume of experimental work falls short of the acceptance standards required by ICLR.

1. The authors have only used their methodology for the correction of wind forecasts, though they have given the reasons of doing this weather forecasts collectively depends on a range of variables and it would be good to test their methodology on other weather variables as well.
2. In experiments MPNN is mostly compared with MLP only (which is bound to improve results), some of the SOTA forecasting models are based on Vision Transformers and Diffusion models which should be explored and added in the comparison here.
3. Reproducibility parameters are missing

1. While the intuition is reasonable, the proposed solution is confused. It seems like a post-calibration method that widely used in the community. For example, instead of doing the ERA5 prediction as proposed in the paper, the prediction results can be further calibrated with the off-grid observations, in this way, the accuracy will be greatly improved. In this case, what is the difference between the proposed approach and the above method. I also would like to see the results.
2. Missing baselines. There are many SOTA baselines for either grid-based predictions or off-grid predictions, but the authors only compare with some basic methods, more baselines are needed to validate the effectiveness of the proposed approach.
3. Since the method rely on the forecasting results of ERA5, the accuracy will be affected heavily by its values. More experiments and discussion should be added.