the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Comparative Analysis of Continuous and Reinitialized Dynamical Downscaling in the North Atlantic and Surrounding Continents
Abstract. General Circulation Models provide comprehensive climate projections but are limited by coarse spatial resolution. To address this issue, Regional Climate Models are used for higher-resolution simulations, particularly to assess regional climate change impacts. This process, called dynamical downscaling, typically involves continuous simulations over a selected period. Alternatively, multiple reinitialized simulations over shorter intervals can be employed to minimize error accumulation and reduce computation time through parallel processing. However, this approach may hinder the development of certain atmospheric phenomena. In this study, the Weather Research and Forecasting model was used for continuous and daily reinitialized dynamical downscaling. Simulations were driven by ERA5 and Coupled Model Intercomparison Project Phase 6 data at 1° and 1.25° spatial resolution, respectively, covering 115° W–40° E in longitude and 20° N–60° N latitude, and downscaled to 20-km resolution. The results were evaluated against ERA5 data at 0.25° resolution to assess accuracy. The analysis focused on wind speed, temperature, humidity, precipitation, surface pressure, and solar radiation. Overall, both techniques demonstrated good to excellent correlation with the reference data. However, neither method reliably captured wind speed nor surface pressure in mountainous areas. In ERA5-driven simulations, the reinitialized technique performed better than the continuous for air temperature and humidity in coastal regions, whereas the continuous approach showed a slight advantage in estimating solar radiation across all surfaces. For CMIP6-driven simulations, both downscaling techniques produced similar results, except for solar radiation and over land, where the continuous method demonstrated marginally better performance. Considering the significantly lower computational cost of the reinitialized method – approximately 30 times less in this study – it is recommended as the preferred approach when its performance is comparable to or better than that of the continuous method.
- Preprint
(3820 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-1339', Anonymous Referee #1, 11 May 2025
This manuscript investigates the performance of continuous versus daily reinitialized dynamical downscaling using the Weather Research and Forecasting (WRF) model. The simulations are driven by ERA5 and CMIP6 data over a large domain covering a wide area in the mid-latitudes of the Northern Hemisphere and are downscaled to 20-km resolution. The study compares both downscaling approaches in terms of their ability to reproduce key atmospheric variables such as wind speed, temperature, humidity, precipitation, surface pressure, and solar radiation, with ERA5 data used as the evaluation reference. The authors aim to assess whether the reinitialized approach—despite its lower computational cost—can produce results comparable to the more resource-intensive continuous simulations. Overall, while the manuscript addresses a relevant topic within regional climate modeling, the current version lacks sufficient novelty and depth of analysis to warrant publication. Key limitations in methodology and interpretation are not adequately addressed, and the conclusions are primarily descriptive without offering new insights into the physical processes or broader implications. Therefore, I do not recommend the manuscript for publication in its current form. To support this recommendation, I outline the main concerns as follows:
- The methodology section lacks novelty and does not present a sufficiently innovative approach to advance the field. The study primarily compares continuous and daily reinitialized dynamical downscaling using the WRF model. The comparison between these two approaches is meaningful, but it does not offer any new insights in methodology. In addition, the experimental design is fairly conventional and lacks a thorough exploration of different experimental setups or configurations. For instance, the study does not investigate the impact of varying reinitialization intervals. Without a more creative or refined experimental design, the manuscript does not sufficiently push the boundaries of current understanding in regional climate modeling.
- The manuscript provides a general description of the results but fails to explore the underlying physical mechanisms behind the observed differences between the two downscaling techniques. While the study acknowledges that neither method reliably captures wind speed and surface pressure in complex terrain, particularly in mountainous regions, it does not sufficiently address the physical processes that contribute to these limitations. Without addressing these physical mechanisms involved, the study does not offer a thorough understanding of the factors influencing the performance of the downscaling techniques.
- The manuscript presents the analysis of various atmospheric variables, but these variables such as wind speed, temperature, humidity, and precipitation are analyzed independently, without considering their interdependencies. In regional climate modeling, it is crucial to explore the interactions between these variables, as they often exhibit complex relationships that can impact model accuracy. The absence of such an integrated analysis limits the depth of the study's findings and fails to offer a better understanding of the underlying climate processes.
Besides, there are several minor issues that need to be addressed.
- Lines 17-18: “115ºW-40ºE in longitude and 20ºN-60ºN latitude” should be “115ºW-40ºE in longitude and 20ºN-60ºN in latitude”.
- Lines 99 and 114: “from -115ºW to 40ºE” should be “from 115ºW to 40ºE” or “from -115ºE to 40ºE”.
- In Figure 2, the latitude and longitude labels are too small and difficult to read. I recommend increasing the font size to improve readability and overall presentation quality.
- Figures 3, 5, 7, 11, 13 and 15: Please check the figure legend — it seems that “renitialized” is a typo, and it should be corrected to “reinitialized”.
Citation: https://doi.org/10.5194/egusphere-2025-1339-RC1 -
RC2: 'Comment on egusphere-2025-1339', Anonymous Referee #2, 08 Aug 2025
This study compares two dynamical downscaling techniques—continuous and daily reinitialized simulations—using the Weather Research and Forecasting (WRF) model. The simulations were driven by ERA5 (upscaled data 1 every 4) and CMIP6 datasets and covered a broad domain. The downscaled data were evaluated against high-resolution ERA5 (all points) data to assess accuracy in key climate variables. The authors present that both methods showed good to excellent agreement with the reference data overall but for wind speed or surface pressure in mountainous areas. From these results the authors state that given that the reinitialized method requires significantly less computational time, approximately 30 times less in this study, it is recommended as the preferred approach since its performance is comparable to or better than that of continuous downscaling.
While the manuscript covers an important subject in regional climate modeling, it does not present enough innovation or thorough analysis in its current form. Critical issues related to the methodology and interpretation of results are not adequately explored, which weakens its overall scientific value. Therefore, I cannot recommend this paper for publication and it should be rejected. Here I expose some reasons.
From the computational point of view, the authors repeatedly fail to clearly distinguish between computational cost and wall-clock time. While the reinitialization approach (i.e., time-parallel integration) can reduce wall time if sufficient computational resources are available, it is inherently more expensive in terms of total computational cost. This is due to the increased number of model spin-ups and the lack of continuity in the simulations. The manuscript should clarify this distinction, as it is critical for properly assessing the efficiency and scalability of the downscaling methods. Note that if, for example, NN CPUs are available and the model configuration allows for their efficient use in parallel (i.e., with minimal loss in scalability), then the computational cost of a continuous simulation would be significantly lower. In such cases, running a single continuous simulation can be more efficient than multiple reinitialized runs, both in terms of total CPU time and resource utilization. This highlights the importance of considering scalability and resource availability when evaluating the computational efficiency of each approach. The message conveyed by the authors may be misleading or even problematic for less experienced users. Without a clear distinction between computational cost and wall-clock time, and without properly addressing the implications of model scalability and resource allocation, there is a risk that users may draw incorrect conclusions about the efficiency of the reinitialized approach. This could lead to inefficient use of computational resources or misinformed methodological choices in future studies.
On the other hand, the authors seem to overlook the fundamental purpose of dynamical downscaling. If the goal is merely to enhance spatial resolution, the time-slice integration approach can be useful, as it acts like a form of pseudo-nudging—allowing the incorporation of higher spatial detail with physical meaning, particularly in regions with complex topography. However, this method can introduce discontinuities in the time series, and its application should be approached with caution, depending on the physical processes being studied. For instance, when used for regional climate change projections, insufficient spin-up time may lead to physical inconsistencies, undermining the reliability of long-term simulations.
if I have understood correctly, the authors compare the output of GCM-driven simulations with reanalysis data just for one year. This is a serious methodological flaw. Reanalysis data represent an aproximation to observed atmospheric states, while GCM-driven simulations for a given year has nothing to do the reality. It is like compare one year to another. Comparing them directly, without proper bias correction and in a climatological context, is not scientifically valid and can lead to misleading conclusions.
The paper is unnecessary long, specially the results section. Probably all results could be reduced to a couple of tables and panels.
The physical interpretation is largely absent, as the authors do not provide any explanation of the underlying physical processes responsible for the differences observed between the downscaling approaches or physical consistence of the atmosphere and surface (for example) fields.
The experimental design is basic and I would say wrong. For example, the authors use a very big domain. Later they use just a part (data above 20°N was selected) because they think the results are worse in tropical-equatorial regions. Also, the skill scores used are not the best for checking one year of data.
Citation: https://doi.org/10.5194/egusphere-2025-1339-RC2
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
603 | 184 | 11 | 798 | 13 | 36 |
- HTML: 603
- PDF: 184
- XML: 11
- Total: 798
- BibTeX: 13
- EndNote: 36
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1