Dear Editor and Reviewers: Thank you very much for your professional and insightful comments concerning our manuscript “ENSO-ASC 1.0.0: ENSO Deep Learning Forecast Model with a Multivariate Air–Sea Coupler”

Thank you very much for your professional and insightful comments concerning our manuscript “ENSO-ASC 1.0.0: ENSO Deep Learning Forecast Model with a Multivariate Air–Sea Coupler” (ID: gmd-2021-213). Those comments are all very valuable and helpful for revising and improving our manuscript, we have made extensive revisions and corrections according to the nice suggestions. The point-by-point major responses are as following for three reviewers respectively. We also edit our manuscript thoroughly for correcting the expression and grammar errors, which is also uploaded as the corresponding track-changes file.


Comment 3:
A suggestion: In this paper, it is found that the best effect is to set the input sequence length as 3. This may be due to selecting the predictors with short memory (vapor, cloud). If predictors with long memory (such as heat content) are added, it may be more effective to set the length longer. Although Table 3 shows the prediction effect of the model with increased heat content data, the input sequence length is the same. This may be taken into consideration in a future study using global data.

Response:
We thank the reviewer for this valuable insight very much. As the reviewer said, from our experiments in Section 4.3.1 (Influence of the input sequence length), we found that the suitable input sequence length for the ENSO-ASC is 3 months according to the trade-off between the time-/resource-consuming and forecast skill when using 6 predictors (SST, u-wind, v-wind, rain, cloud, and vapor), in which 5 variables are related to the atmospheric processes with short memories. In the Section 4.4.1 (Contributions of different predictors to the forecast skill), when we add heat content with long memory into the model, it is indeed necessary to re-investigate the optimal input sequence length by experiments in this manuscript. In fact, in continuous studies following this manuscript, the input length should be at least 6 months with 7 input variables (SST, u-wind, vwind, rain, cloud, vapor and heat content) using globe data. While with the equatorial Pacific data and the input sequence length varying from 3 to 9 months, the change of forecast skill of ENSO-ASC is not much significant. Because the input region mainly covering the equatorial Pacific and most of the variables are with short memories in this manuscript, the input sequence length is still set as 3 despite adding heat content data into the model shown in Table 3. Let's look forward to our next manuscript following this manuscript.
The related statements are additionally supplemented in the Section 4.4.2 (Contributions of different predictors to the forecast skill) at line 500 as the blue text below: "The superiority of our proposed model derives from the graph formalization, and the special multivariate coupler can effectively express the processes of synergies between multi-physical variables.
From another perspective, the improvement of the forecast skill is not only benefited from graph formalization, but also due to the utilization of multiple variables highly related to ENSO compared to using limited variable to predict ENSO as previous works. For ENSO forecast, SST is definitely the most critical predictor. Besides SST, other variables have different contributions to the forecast results. Therefore, we design an ablation experiment by removing one of predictors from our proposed model and detect the reduction of forecast skill (Table 3 above). At the meanwhile, we also add one extra predictor (from surface air temperature, surface pressure and ocean heat content respectively) into our proposed model to investigate the improvement of forecast skill (Table 3 below). Here, the input sequence length is still set to 3.   It is worth noting that the input sequence length should be longer when feeding the ocean heat content into the multivariate coupler, because this predictor is with long memory. However, as the input sequence length varies from 3 to 9 months, the forecast skills of ENSO-ASC have not changed much actually. This is mainly because that the global spatial teleconnections and temporal lagged correlations by Walker Circulation and ocean waves

(such as Kelvin and Rossby Waves) (Exarchou et al., 2021 and Dommenget et al., 2006) are not caught in the model, the input region of which mainly covers the equatorial Pacific. In addition, the model contains only one long memory predictor besides SST.
In the subsequent experiments, the model will use the chosen 6 variables (SST, u-wind, v-wind, rain, cloud, and vapor)

and the input sequence length is set to 3."
The related statements are also additionally supplemented in the Section 5 (Discussions and conclusions) at the line 845 as the blue text below:

"The extensive experiments demonstrate that the ENSO forecast model with a multivariate airsea coupler (ENSO-ASC) is a powerful tool for analysis of ENSO-related complex mechanisms.
Meteorological research does not only pursue skilful models and accurate forecasts, but requires a comprehensive understanding of the potential dynamical mechanisms. In the future, we will extend our model to more global physical variables with informative vertical layers, such as the thermocline depth, and the ocean temperature heat content, to explore the global spatial remote teleconnections, temporal lagged correlations, and the optimal precursor etc." The related references are shown as following and also added into the manuscript: References Dommenget, D., Semenov, V., and Latif, M.: Impacts of the tropical Indian and Atlantic Oceans on ENSO, Geophysical research letters, 33,2006. Exarchou, E., Ortega, P., Rodríguez-Fonseca, B., Losada, T., Polo, I., and Prodhomme, C.: Impact of equatorial Atlantic variability on ENSO predictive skill, Nature communications, 12, 1-8, 2021.

The Reply on Referee #2
Comment 1: In the ablation experiment, "The calculation of this variable contains SST, so the effect of the extra introduction of upper ocean heat content will be weakened" is at L533. I have a suggestion: if using upper ocean heat content to take place the SST in the model, how will the ENSO-ASC perform?
Response: Thank you so much for your professional attitude and insightful suggestion. This is indeed a valuable question for investigating the effects of different predictors on an ENSO deep learning forecast model. The upper ocean heat content is a very concerned variable, which can reflect the vertical and horizontal propagations of ocean waves and help interpret the dynamical mechanisms of ENSO. Therefore, as the comment says, we supplement a control experiments to investigate the model performance by replacing SST with upper ocean heat content in the model input.
We conduct the comparison by two modified ENSO-ASCs with the same output of SST + uwind, v-wind, rain, cloud, and vapor, while with the different input. One is upper ocean heat content + u-wind, v-wind, rain, cloud, and vapor (EXAM), the other is SST + u-wind, v-wind, rain, cloud, and vapor (CTRL). We find that the forecast skill of EXAM is slightly lower than that of CTRL (depicted as Table 4). The upper ocean heat content is the average of the oceanic temperature from sea surface to upper 300m, which is crucial to represent the deeper sea temperature beyond sea surface. However, our model is designed to forecast SST. We think that using the upper ocean heat content as a predictor for our model inevitably introduces more noise, which extracts the features of oceanic temperature not only from sea surface but also from deeper ocean. Actually, according to our extensive experiments, we find it is a positive determination that the model should select the physical variable we want to forecast as one of predictors.
We also supplement the related statements from the start of line 545 as the blue text below: "Among the three extra added physical variables, the upper ocean heat content is a very concerned variable, which can reflect the vertical and horizontal propagations of ocean waves and help interpret the dynamical mechanisms. Therefore, we conduct the comparison via two modified ENSO-ASCs with the same output of SST + u-wind, v-wind, rain, cloud, and vapor, while with the different input. One uses upper ocean heat content + u-wind, v-wind, rain, cloud, and vapor, marked as EXAM, another uses SST + u-wind, v-wind, rain, cloud, and vapor, marked as CTRL. The results are shown in Table 4. Note: Model paradigm represents the input and the output for the ENSO-ASC, where → means "forecast". "Others" is five variables, including u-wind, v-wind, rain, cloud, and vapor. The first row is the control experiment, which is the same with the result in Table 3, and the second row is the examined experiment, which only replaces SST by upper ocean heat content in the model input.
The forecast skill of EXAM is slightly lower than CTRL. The upper ocean heat content is the average of the oceanic temperature from sea surface to upper 300m. When using it as a predictor to forecast SST, our model will extract the features of oceanic temperature not only from sea surface but also from deeper ocean, which inevitably introduces more noise. This may be a reason for the above result. Therefore, we still use SST instead of the upper ocean heat content as the key predictor which would bring higher forecast skill." Comment 2: The initial letter of the sentence should be uppercase and some mistakes are found at line 103 and line 222.