Predicting Sea Surface Temperatures with Coupled Reservoir Computers
- 1Department of Physics and Department of Mechanical Engineering, Clarkson University
- 2Department of Electrical and Computer Engineering, Clarkson University
- 3Clarkson Center for Complex Systems Science
- 1Department of Physics and Department of Mechanical Engineering, Clarkson University
- 2Department of Electrical and Computer Engineering, Clarkson University
- 3Clarkson Center for Complex Systems Science
Abstract. Sea surface temperature (SST) is a key factor in understanding the greater climate of the Earth and is an important variable when making weather predictions. Methods of machine learning have become ever more present and important in data-driven science and engineering including in important areas for Earth Science. We propose here an efficient framework that allows us to make global SST forecasts by use of a coupled reservoir computer method that we have specialized to this domain allowing for template regions that accommodate irregular coastlines. Reservoir computing is an especially good method for forecasting spatiotemporally complex dynamical systems, as it is a machine learning method that despite many randomly selected weights, it is nonetheless highly accurate and easy to train. Our approach provides the benefit of a simple and computationally efficient model that is able to predict sea surface temperatures across the entire Earth’s oceans. The results are demonstrated to replicate the actual dynamics of the system over a forecasting period of several weeks.
Benjamin Walleshauser and Erik Bollt
Status: final response (author comments only)
-
RC1: 'Comment on npg-2022-4', Anonymous Referee #1, 02 Mar 2022
The draft manuscript titled “Predicting Sea Surface Temperatures with Coupled Reservoir Computers” is an excellent effort in using coupled reservoir computers for predicting global sea surface temperatures. The manuscript may be accepted after minor revision after the following comments have been addressed:
- Line 125 says that the actual values of SST were used for training. Can the authors comment on why normalization was not used as it has been shown to be necessary to train machine learning models?
- The authors train on a daily SST dataset, now oceans are known to operate on long temporal scales having a memory of at least a month. Can the authors comment on the utility of these forecasts? For example, Nino3, or Nino3.4 are considered based on monthly datasets precisely for the reason that the oceanic processes are slow.
- Was any hyperparameter tuning performed?
- From figure 4, it can be seen that the model performs well where there is an established trend. For example, Fig 4a-f have a clear trend and the model is performing very well in all of them. There are some deviations in Fig 4g whereas Fig 4h is showing good performance. The intent of using such as coupled reservoir computer is to simulate the chaoticity of the system, whereas out of the 8 subplots in Figure 4, 6 have a clear linear trend where the model performs well, whereas in Fig 4g where there are some deviations, the model is not performing good relative to the previous subplots. Fig 4h is satisfactory. Can the authors maybe provide some more examples or describe these features from the results?
- AC1: 'Reply on RC1', Ben Walleshauser, 08 Apr 2022
-
RC2: 'Comment on npg-2022-4', Anonymous Referee #2, 10 Mar 2022
A domain-decomposed set of coupled reservoirs
that share the same set of hyperparameters are trained over multiple
years of reanalysis daily sea surface temperature (SST) data and used to
predict SST at lead times of between one day and six weeks.The application of RC to the specific problem the authors consider,
the analysis of the results and the write-up, all seem to be of a
somewhat preliminary nature. As such it is not clear what a potential
reader is expected to take away from this article. This issue needs to
be addressed in a substantive fashion to be further considered for
publication. A few other issues are noted below.The authors state: To observe the effect of
the randomly selected input and middle weights on the performance of
the RCs, the model was ran 15 times all with the same metaparameters
as described in Table 1, to collect data for the examination of the
error.
Please state how the ensemble of predictions that use random
variations of input and middle weights was analyzed. Meaning, what
error is being shown is say figures 4-13
In the context of Figs. 6 and 7, while the authors chose to
spinup/warmup the reservoirs over a period of a week, the results
suggest that a longer spinup of the reservoirs (of about 4 weeks) is
called for. Please comment on the a priori choice of one week and the
longer timescale that is required as indicated by the results. How
does the longer timescale vary with changes in the hyperparameters?
What is the relevance of the leakage parameter in the context of a
leaky reservoir update in this context? (and which the authors do not consider)Please comment on possible reasons for 4-5 day timescale seen in
Figure 8, particularly since the data itself, e.g., as seen in fig. 4, seems
to display variations on a broader range of timescales.Given the results presented, it may seem somewhat of an
over-statement when the authors state that "The results are
demonstrated to replicate the actual dynamics of the system over a
forecasting period of several weeks."- AC2: 'Reply on RC2', Ben Walleshauser, 08 Apr 2022
Benjamin Walleshauser and Erik Bollt
Benjamin Walleshauser and Erik Bollt
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
500 | 81 | 11 | 592 | 4 | 2 |
- HTML: 500
- PDF: 81
- XML: 11
- Total: 592
- BibTeX: 4
- EndNote: 2
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1