the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
16 Feb 2021
16 Feb 2021
Producing realistic climate data with GANs
- 1CERFACS, Toulouse, France
- 2CNRM, Université de Toulouse, Météo-France, CNRS, Toulouse, France
- 3Institut de Mécanique des Fluides de Toulouse (IMFT), Université de Toulouse, CNRS, Toulouse, France
- 4Institut National Polytechnique de Toulouse
- 1CERFACS, Toulouse, France
- 2CNRM, Université de Toulouse, Météo-France, CNRS, Toulouse, France
- 3Institut de Mécanique des Fluides de Toulouse (IMFT), Université de Toulouse, CNRS, Toulouse, France
- 4Institut National Polytechnique de Toulouse
Abstract. This paper investigates the potential of a Wasserstein Generative Adversarial Networks to produce realistic weather situations when trained from the climate of a general circulation model (GCM). To do so, a convolutional neural network architecture is proposed for the generator and trained on a synthetic climate database, computed using a simple 3 dimensional climate model: PLASIM.
The generator transforms a latent space
, defined by a 64 dimensional Gaussian distribution, into spatially defined anomalies on the same output grid as PLASIM. The analysis of the statistics in the leading empirical orthogonal functions shows that the generator is able to reproduce many aspects of the multivariate distribution of the synthetic climate. Moreover, generated states reproduce the leading geostrophic balance present in the atmosphere.
The ability to represent the climate state in a compact, dense and potentially nonlinear latent space opens new perspectives in the analysis and the handling of the climate. This contribution discusses the exploration of the extremes close to a given state and how to connect two realistic weather situations with this approach.
Camille Besombes et al.
Status: final response (author comments only)
-
RC1: 'Comment on npg-2021-6', Anonymous Referee #1, 17 Mar 2021
This paper introduces a novel method of producing realistic climate data using Generative Adversarial Networks. The paper is well written in general. The text is fluent. The scientific approach is sound and clear. The results are well presented.
Few questions.
1. How tolerant are the wrap padding layers to discontinuity? To which extent, the padding process may affect the training?2. By looking at the inputs/outputs in figures 1 to 5, I assume that some convolutional blocks have pooling/ upsampling layers and others do not. Could you explain the reason for that and present these blocks differently in the figures?
3. Figure 2 and 3 present two different architectures of the residual blocks of the Critic. Which one was used?
4. The stopping criteria of the training is a maximum number of iterations. Could this condition be combined with the convergence of the loss function, for example loss function derivative very small or training set and validation set having almost equal loss?
5. I would suggest reducing the length of the paper as follows:
1. figure 10 and 11 could be merged in one figure
2. figure 12 and 13 present almost the same information. I would recommend keeping only one of them.Minor comments.
1. Please review the formatting of equations 5, 10, 13
2. Figure 2 and 3 have the same title
3. Please use the same title pattern in figure 5 as in figure 2
4. Figure 21, 22 and 24, please review the titles and explain what is a, b, c, d
5. Figure 29, please use the equation reference, (as in 28) instead of the equation itself. -
RC2: 'Comment on npg-2021-6', Anonymous Referee #2, 25 Apr 2021
I apologize for the delay in submitting the report on the manuscript submitted to Nonlinear Processes in Geophysics.
The manuscript developed a weather generator with a convolutional neural network and investigated its performance. The manuscript shows that the generator succeeded in simulating the atmosphere in many regards. The manuscript is exceptionally well written, and the topic would attract many readers of the Nonlinear Processes in Geophysics. I only have a few comments listed below. I recommend that the manuscript is accepted for publication after all the points are properly mentioned.
I feel that the manuscript can attract more readers by providing more information on the background of the study. How about adding more references about weather generators in the introduction?
The authors evaluated the spatial distribution of the variables simulated by the generator and showed that the generator’s performance is generally good. This is good to evaluate the generator but lacks thoroughness. The authors should evaluate how well the inter-variable relationships are simulated by the generator. It is also better to investigate the reproducibility of the vertical structure.
L197: is it possible to generate weathers for a specific date? Similarly, is it possible to give the generator SST fields (or any other boundary conditions (e.g. GHGs, orbital parameters, solar, etc.)) as inputs? For the generator to be used for applications you mentioned in the manuscript (e.g. risk assessment, data assimilation), such information should be included as the inputs.
The most important advantage to use weather generators is, I believe, its low computational cost. Therefore, the authors should show how fast the weather generator is compared to the GCM used as the training data.
Camille Besombes et al.
Model code and software
Producing-realistic-climate-data-with-GANs Camille Besombes https://doi.org/10.5281/zenodo.4442450
Camille Besombes et al.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 555 | 88 | 7 | 650 | 4 | 4 |
- HTML: 555
- PDF: 88
- XML: 7
- Total: 650
- BibTeX: 4
- EndNote: 4
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1