Enhancing 1-month temperature predictions in South Korea through dynamical downscaling of machine learning global ensemble forecasts

Enhancing 1-month temperature predictions in South Korea through dynamical downscaling of machine learning global ensemble forecasts | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Enhancing 1-month temperature predictions in South Korea through dynamical downscaling of machine learning global ensemble forecasts Subin Ha, Xiaohui Zhong, Lei Chen, Hao Li, Jina Hur, Hyun-Han Kwon, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7639295/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract Despite the growing availability of global ensemble forecasts, their application at regional scales remains challenging due to the coarse spatial resolution. Meanwhile, machine learning (ML)-based forecasts are emerging as promising alternatives to traditional physics-based models, yet they are mostly deterministic, and their potential has not been fully explored beyond the medium-range timeframe. This study investigates the potential of integrating global ML-based ensemble forecasts (FuXi-ENS) with dynamical downscaling to improve one-month temperature prediction over South Korea. The forecasting performance of FuXi-ENS, in terms of both temporal and spatial patterns, is compared against state-of-the-art physics-based model forecasting data from NOAA and ECMWF, which serve as benchmarks. The superiority of FuXi-ENS becomes pronounced after dynamical downscaling, highlighting the added value of ML-based forecasts when combined with high-resolution physical modeling. Overall, this study offers a comprehensive assessment of extended-range temperature prediction over South Korea, illustrating the operational potential of hybrid approaches that combine global ML models with regional dynamical downscaling and providing insights for the future development of hybrid forecasting systems. Earth and environmental sciences/Climate sciences Earth and environmental sciences/Environmental sciences Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Subseasonal-to-seasonal (S2S) forecasting plays an increasingly important role in bridging the critical gap between short-term weather forecasts and long-term seasonal outlooks 1 – 4 . Accurate forecasts at this timescale are essential for effective resource planning and risk management in various sectors such as agriculture, water resource management, energy, public health, and disaster preparedness, where decisions often depend on climate conditions weeks to months in advance 5 – 7 . In response to this need, global seasonal forecasting systems have been developed and extensively assessed across regions and variables, ranging from near-surface meteorological variables like temperature and precipitation to large-scale climate phenomena such as El Niño and seasonal monsoons 8 – 20 . For South Korea, several studies have assessed the performance and utility of various seasonal forecasts 21 – 25 . Using the Climate Forecast System version 2 (CFSv2) operational forecasts, Ha et al. 21 found that forecasts with shorter lead times are more likely to perform better in capturing interannual temperature variability of South Korea. Still, the relatively coarse spatial resolution of global forecasts remains a major obstacle, limiting their ability to represent localized variability and hindering their direct application for end users. To address these shortcomings, dynamical downscaling has been utilized by a few studies to enhance regional details through high-resolution simulations. Im et al. 22 and Oh et al. 25 showed that applying dynamical downscaling to CFSv2 forecasts significantly improves their temperature prediction for the country, enabling better use of forecasts for downstream applications such as calculating temperature-based agricultural indices and predicting heatwaves. These improvements highlight the added value of dynamical downscaling of global forecasts. While physical models have been the mainstay of global forecasting, machine learning (ML) has rapidly led to significant advances in forecasting across different timescales, including the S2S range. Recent studies have demonstrated that ML-based forecasts can match or surpass the performance of traditional physical models. In particular, a number of studies have demonstrated that the FuXi-ENS model outperforms the European Centre for Medium-Range Weather Forecasts (ECMWF)’s state-of-the-art system 26 – 30 . Other ML approaches have also yielded improved temperature and precipitation forecasts in various regions, including China, Brazil, West Africa, and the U.S., consistently outperforming traditional physics-based models for subseasonal timescales 31 – 35 . Despite these promising advances, applications of ML-based forecasting in South Korea remain scarce, and existing work is mostly limited to nowcasting or short-term prediction 36 – 38 . As ML forecasting systems continue to advance, there is a pressing need to evaluate their performance in South Korea and explore ways to enhance their utility by integrating techniques such as dynamical downscaling in order to bridge the scaling gap between prediction and actionable information. In this regard, this study evaluates the potential of combining ML-based global forecasts with dynamical downscaling to improve 1-month temperature prediction over South Korea, with a focus on the month of July from 2018 to 2024. Specifically, we assess and compare the performance of three global forecasts: the Seasonal Forecasting System version 5 (SEAS5) from ECMWF, the CFSv2 from the National Centers for Environmental Prediction (NCEP), and the ML-based FuXi-ENS model developed by Fudan University. The forecast skill of each system is first evaluated against observed daily temperature data in South Korea. To examine the added value of high-resolution simulations, dynamical downscaling is then applied to selected ensemble members from CFSv2 and FuXi. The downscaled outputs are subsequently evaluated against observations. As far as our knowledge extends, this is the first attempt to apply dynamical downscaling to an ML-based global forecast system to obtain fine-scale forecasting information with a 1-month lead time targeted at South Korea. Moreover, by focusing on daily temperature forecasts rather than monthly averages, this study aligns more closely with the needs of end users, who often require high-frequency, locally actionable information. This comparative assessment will provide valuable insights into the feasibility and effectiveness of integrating readily available global forecasts and a customized dynamical downscaling system for the target region for improving subseasonal forecasting, ultimately contributing to more informed and timely decision-making in climate-sensitive sectors. Methods Study domain and observed data (OBS) This study focuses on one-month temperature prediction on a daily timescale for July during 2018–2024 in South Korea. OBS for this period were obtained from the Korean Meteorological Administration. Each station measures 2-m air temperature at hourly intervals, and the daily mean temperature (Tmean) was computed as the average of the 24 values. Tmean records from 506 observation stations were used in this study, including 88 Automated Synoptic Observing System (ASOS) stations and 418 Automatic Weather System (AWS) stations. The spatial distribution of these stations is shown in Fig. S1 online. Global seasonal forecasts Three global seasonal forecast models are employed in this study: SEAS5, CFSv2, and FuXi-ENS. SEAS5 39 is the latest seasonal forecasting system developed by ECMWF, consisting of a 51-member ensemble. Operational since November 2017, SEAS5 is initialized on the 1st of each month and provides forecasts extending up to 7 months. Surface-level data are available at 6-hour intervals, while pressure-level data are available at 12-hour intervals. Due to the coarse temporal resolution of upper-level fields, SEAS5 cannot be directly used as input for dynamical downscaling. Instead, 6-hourly 2-meter temperature data at a horizontal resolution of 1° are employed as a benchmark for selecting ensemble members from CFSv2 and FuXi-ENS for dynamical downscaling. Daily Tmean from SEAS5 is derived from the four available 6-hourly values each day. CFSv2 40 operational forecasts by NCEP have been available since April 2011 and are publicly available through the NCEP data archive. The forecasting system is initialized four times daily (00, 06, 12, and 18 UTC) and produces deterministic forecasts up to nine months at 6-hour intervals. Surface-level forecast data are available at approximately 100 km resolution (Gaussian T126), while pressure-level data are provided at 0.5° resolution. For this study, 40 ensemble members initialized between June 21st UTC00 and June 30th UTC18 for each year are used to generate July forecasts. Daily Tmean is obtained by averaging the four 6-hourly forecasts available per day. FuXi-ENS 30 is a machine learning-based seasonal forecasting system developed using ERA5 as training data, with a spatial resolution of 0.25°. Trained on ERA5 data through 2017, FuXi-ENS produces forecasts available from 2018 onward. While originally designed for 0–15 day forecasts, here FuXi-ENS is extended to 33 days to assess its one-month prediction skill. Each year, 40 perturbed ensemble members were initialized at 00 UTC on June 29th, with each forecast providing 6-hourly data at both surface and pressure levels through the end of July. Like the other two aforementioned models, daily Tmean from FuXi-ENS is constructed as the arithmetic mean of its four 6-hourly outputs per day. Study period and ensemble member selection for dynamical downscaling Both CFSv2 and FuXi-ENS provide 40 ensemble members, respectively, for July forecasts each year. Due to computational constraints, it is not feasible to dynamically downscale all ensemble members over the entire study period. Therefore, 4 representative members (10% of the ensemble) are selected from each forecast system per year for dynamical downscaling. For historical periods, observed temperature records are available, so ensemble members can be selected based on their agreement with OBS before downscaling. However, the ultimate goal of this research is to develop an operational system for producing practical 1-month forecasts for South Korea, applicable to future periods when OBS are not yet available. Thus, for a generalizable and operationally feasible member selection strategy, a proxy reference is required. SEAS5 is used as this benchmark based on its generally superior forecasting skills, as demonstrated in previous studies 16 , 41 . Figure 1 presents the correlation coefficients and root-mean-square errors (RMSEs) of daily Tmean between OBS and forecasts in South Korea from each global model for July in each year. In most years, SEAS5 outperforms CFSv2 and FuXi-ENS, showing higher correlation coefficients and smaller RMSEs against OBS. Furthermore, despite having more ensemble members (51) than CFSv2 and FuXi-ENS (40), SEAS5 generally exhibits smaller inter-member variability, as indicated by narrower boxplots for both correlation and RMSE. While these results generally support the use of SEAS5 as a proxy for observed conditions in member selection since 2018, when FuXi-ENS forecasts became available, the dynamical downscaling of CFSv2 and FuXi-ENS for comparative assessment is conducted for the years 2018, 2019, and 2020, when there is no missing data for CFSv2. Additionally, these three consecutive years represent a distinct range of July temperature conditions: in the recent 20 years (mean climatology: 24.6°C), 2018 experienced the hottest July on record (mean temperature: 26.6°C), 2020 the coolest (mean temperature: 22.5°C), and 2019 conditions close to the climatological average (mean temperature: 24.7°C). Model performance also varied markedly among these years. In 2018, the lower quartile of SEAS5 members’ correlation coefficients exceeded 0.8, indicating consistently strong agreement with OBS across nearly the entire ensemble members; FuXi-ENS members also showed high correlations. In contrast, in 2020, even the upper quartile of SEAS5 correlations was only ~ 0.25, highlighting particularly weak forecast performance; CFSv2 and FuXi-ENS exhibited similarly low correlations. Interestingly, RMSEs in 2020 were small for all three models, with narrow inter-member ranges. This indicates that while the models fail to capture daily fluctuations, their absolute errors are relatively small. The year 2019 displayed intermediate characteristics. These three years, therefore, represent cases with distinct observed temperature conditions and differing forecast skill across the three systems, warranting detailed analysis in the following sections. The selection process proceeds as follows: For each year, daily Tmean values of the 51 SEAS5 members are averaged to form the ensemble mean, hereafter referred to as SEAS5-ENS. Then, the daily Tmean of each CFSv2 and FuXi-ENS ensemble member is compared to SEAS5-ENS by calculating the Pearson correlation coefficient for July. The 4 members from each model with the highest correlations to SEAS5-ENS are selected for dynamical downscaling. Model configuration for dynamical downscaling The Weather and Research Forecasting (WRF) model version 4.5 is employed for dynamical downscaling of CFSv2 and FuXi-ENS forecasts. A one-way nested domain setup is used, consisting of a mother domain at 20 km resolution and a nested domain at 5 km resolution, both defined using a Lambert conformal projection centered at 38°N and 127.5°E (see Supplementary Fig. S2 online). The physical parameterization schemes follow the optimal configuration identified by Qiu et al. 42 through sensitivity testing. Specifically, the model employs the WSM3 microphysics scheme 43 , the RRTMG longwave and shortwave radiation schemes 44 , the Revised MM5 surface layer scheme 45 , the Noah land surface model 46 , the Yonsei University planetary boundary layer scheme 47 , and the Kain–Fritsch cumulus convection scheme 48 with trigger function of Ma and Tan 49 . Successful initialization of WRF requires a minimum set of meteorological and land-surface fields. While CFSv2 forecasts provide all necessary variables, FuXi-ENS lacks several key fields, including skin temperature, surface pressure, soil moisture, and soil temperature. To address this gap, ERA5 reanalysis data are incorporated. Rather than prescribing daily variability, we construct July climatologies by averaging ERA5 data from 2015 to 2024 at four synoptic times (00, 06, 12, and 18 UTC) to align with the forecast intervals of CFSv2 and FuXi-ENS. This approach assumes that the predictability of FuXi-ENS downscaling primarily derives from the large-scale patterns, while acknowledging that day-to-day variations in surface variables may not be fully captured. Results Skill evaluation of SEAS5, CFSv2, and FuXi-ENS The correlation between Tmean from each ensemble member of SEAS5, CFSv2, and FuXi-ENS with OBS for the study period is shown in Fig. 2 as heatmaps. Consistent with Fig. 1 , the superior performance of SEAS5 is evident, with a dominance of strong red colors indicating high positive correlations. While most members of CFSv2 and FuXi-ENS also exhibit positive correlations with OBS, their intensities are generally weaker, shown by lighter red shades. Notably, a considerable number of members, particularly from CFSv2, are shown in white or near-white, indicating low correlation, and some even appear in blue, highlighting negative correlations with OBS. When comparing CFSv2 and FuXi-ENS, FuXi-ENS appears to slightly outperform CFSv2 overall. However, it's important to note a key distinction in initialization: CFSv2 members are initialized every 6 hours, resulting in different lead times ranging from 240 hours to 6 hours, whereas FuXi-ENS ensemble members are all initialized at 00 UTC on June 29, corresponding to a fixed lead time of 48 hours, with perturbations used to generate ensemble spread. In Fig. 2 , for each year, CFSv2 members are ordered from top to bottom based on lead time, with the longest lead time (ddhh: 2100) on the top and the shortest (ddhh: 3018) on the bottom. It is clearly seen that members with shorter lead times tend to show stronger positive correlations with OBS. This pattern aligns with findings from Ha et al. 21 , which showed that shorter lead times generally yield better temperature forecast skill. On a year-by-year basis, 2018 is especially notable, as all 51 SEAS5 members exhibit very strong positive correlations with OBS. Similarly, in 2019, most SEAS5 members are also strongly correlated. For these two years, FuXi-ENS ensemble members are predominantly positively correlated, and CFSv2 members with short lead times also show good correlation. In contrast, 2020 presents the weakest performance, with most SEAS5 members showing only weak positive correlations, and CFSv2 and FuXi-ENS members showing minimal skill compared to the previous two years. Member selection for dynamical downscaling Following the member selection procedure outlined in Section 2.3.1, SEAS5-ENS is used to compute daily Tmean correlation coefficients with each CFSv2 and FuXi-ENS member. The four members from each system with the highest correlation to SEAS5-ENS are selected for dynamical downscaling and are indicated by green boxes in Fig. 2 . The details of selected members, including initialization times for CFSv2 and member numbers for FuXi-ENS, are provided in Supplementary Table S1 online. To evaluate the selection method, the actual correlation of the selected members with OBS is assessed, and each selected member’s ranking in terms of correlation with OBS among the 40 total members is also provided in Supplementary Table S1 online to offer insight into the relative skill of selected and non-selected members. The effectiveness of the selection method is found to be highly dependent on the predictive skill of SEAS5, which is used as the selection benchmark. In 2018, for example, the selected CFSv2 and FuXi-ENS members—based on correlation with SEAS5-ENS—also turn out to be the top 4 members most highly correlated with OBS. For FuXi-ENS, the ranking of selected members based on SEAS5-ENS matches exactly with their ranking against OBS (1st to 4th), demonstrating a perfect selection outcome. This success reflects the high skill of SEAS5 forecasts in 2018, where all 51 SEAS5 members showed strong positive correlations with OBS. Consequently, SEAS5 served as a reliable reference for guiding member selection. Similarly, in 2019, when SEAS5 forecasts were also skillful, FuXi-ENS member selection performed well, successfully identifying the top 4 OBS-correlated members. The CFSv2 selection in 2019 was also reasonable, capturing the members ranked 3rd, 5th, 6th, and 7th in correlation with OBS. However, in 2020, when SEAS5 members had the weakest correlations with OBS, the selected members from both CFSv2 and FuXi-ENS mostly ranked in the bottom 50% in terms of their actual correlation with OBS, highlighting the challenge of making accurate selections when the benchmark itself lacks predictive skill. Added value of dynamically downscaled forecasts The selected CFSv2 and FuXi-ENS ensemble members were dynamically downscaled using the WRF model configuration described in Section 2.3.2. The mean of the four selected original CFSv2 members is referred to as CFS-OG-ENS4 (OG: original), and the mean of their dynamically downscaled outputs is referred to as CFS-DS-ENS4 (DS: downscaled). Similarly, the FuXi-based original and downscaled ensemble means are denoted as FuXi-OG-ENS4 and FuXi-DS-ENS4, respectively. To evaluate the added value of dynamical downscaling, we first assess the model performance in simulating regionally distinct temperature patterns. Figure 3 shows the spatial distribution of weekly Tmean for OBS, SEAS5-ENS, CFS-OG-ENS4, CFS-DS-ENS4, FuXi-OG-ENS4, and FuXi-DS-ENS4 for July in 2018 and 2020; results for July 2019 are provided in Supplementary Fig. S3 online. In July 2018, temperatures increased steadily from week 1 to week 4. In OBS, the spatial mean was around 22°C in week 1, rose sharply to almost 27°C in week 2, and remained high at 28.1°C and 28.6°C in weeks 3 and 4, respectively. This warming trend was generally reproduced by SEAS5, CFSv2, and FuXi-ENS, although all three underestimated absolute temperatures due to their coarse resolution. SEAS5 and CFSv2, in particular, simulated relatively uniform temperatures across the country, which primarily follow latitudinal gradients, and failed to capture topographically driven variations. In contrast, FuXi-ENS better differentiated regional temperature patterns and exhibited the smallest cold bias, likely due to its comparatively higher native resolution. Dynamical downscaling further improved FuXi-ENS’s spatial representation, capturing colder conditions over mountainous areas and warmer conditions over lowland plains more accurately. Although the spatial mean temperature decreased slightly after downscaling due to the better-resolved cool mountain areas, the enhanced spatial detail provides valuable regional-scale information beyond that available in the original forecasts. CFSv2 also shows enhanced spatial details after downscaling, but its large cold bias in the original forecasts persisted. In weeks 3 and 4, when OBS exceeded 28°C, CFS-DS-ENS4 remained more than 3.5°C cooler than OBS, whereas FuXi-DS-ENS4 was approximately 2°C cooler. In July 2020, observed conditions did not follow a steady warming trend. OBS recorded 22.7°C in week 1, followed by a drop of about 2°C in week 2, a return to week-1 levels in week 3, and only a slight increase (0.5°C) in week 4. This fluctuating pattern was not captured by any of the three original forecast systems, all of which predicted a monotonic week-to-week warming. This suggests that while forecast systems can predict steadily warming conditions (e.g., 2018), they may be less effective at capturing complex temporal fluctuations such as those observed in 2020. After downscaling, FuXi-DS-ENS4 showed a warm bias of only 0.8°C in week 1 but failed to capture the sharp week-2 cooling, increasing the warm bias to 4.6°C, and in weeks 3 and 4, FuXi-DS-ENS4 maintained a warm bias of ~ 3°C. Similarly, while CFS-DS-ENS4 matched OBS well in week 1, it failed to capture subsequent fluctuations, maintaining a warm bias thereafter. Thus, while spatial details improve after dynamical downscaling, the models struggled to reproduce weekly temperature variability, reflecting that the primary source of predictive skill resides in the original large-scale forecasts. In addition to the evaluation of spatial temperature patterns, we further assessed the forecasts in terms of their temporal consistency and spatial fidelity against observations. Specifically, for each of the four ensemble members that constitute ENS4, we computed the temporal correlation of daily Tmean throughout July with OBS and the spatial correlation of monthly Tmean with OBS. Figure 4 presents daily Tmean temporal correlation on the x-axis and monthly Tmean spatial correlation on the y-axis for each ensemble member of CFSv2 and FuXi-ENS. As already demonstrated in the preceding results, the year 2018 exhibited markedly higher forecast skill compared to 2020, a finding that is again confirmed in this figure. In 2018, both CFSv2 and FuXi-ENS members achieved substantially higher temporal correlations, clustering toward the far right side of the x-axis. Notably, CFSv2 benefitted substantially from dynamical downscaling. While temporal correlations remained largely unchanged before and after downscaling, spatial correlations increased from around 0.5 to 0.8. This improvement may be attributed to the refinement from the coarse 1° grid of the original forecasts to the 5 km grid of the downscaled products, which allows each observational station to be represented by values that more closely match local conditions. A similar benefit was found in 2020, with CFSv2 ensemble members improving approximately from 0.4 to 0.8 in spatial correlation after downscaling, while maintaining similar levels of temporal correlation. At first glance, FuXi-ENS appears to gain relatively little from dynamical downscaling, since the increases in station-based spatial correlation are modest. However, the correlation metrics in Fig. 4 are calculated only at available station locations, whereas the spatial maps shown earlier (i.e., Fig. 3 ) clearly demonstrate that FuXi-ENS benefits significantly from dynamical downscaling in terms of improved representation of spatially distinct temperature features. It should also be noted that FuXi-ENS, which already possesses relatively high native resolution, exhibited much higher correlations even before downscaling, with its original forecasts already achieving levels comparable to the downscaled CFSv2 products. Nonetheless, after downscaling, FuXi-ENS still exhibited modest but consistent spatial gains, particularly in 2020 when the original forecasts performed less well, highlighting the added value of higher-resolution simulations. Although a slight decrease in temporal correlation was observed for some FuXi-ENS members, the magnitude of this reduction was rather small and does not materially affect the overall temporal fidelity of the downscaled products. Overall, these results emphasize that while the general predictability of temperature variability originates from the global forecasts themselves, dynamical downscaling enhances the realism of regional-scale representation. The downscaled products inherit the temporal patterns of the original forecasts but resolve local heterogeneity more accurately, producing forecasts that are more realistic and locally relevant. This highlights the complementary role of dynamical downscaling in bridging the gap between global predictability and regional usability. Discussion This study evaluated the performance of 1-month temperature forecasts for July over South Korea using three global forecast systems, SEAS5, CFSv2, and FuXi-ENS, focusing on the years 2018, 2019, and 2020, which represent contrasting temperature patterns. Comparisons with observations from 506 ground-based stations show that SEAS5 consistently exhibits the highest correlation with daily Tmean, making it a practical benchmark for ensemble member selection and subsequent dynamical downscaling of CFSv2 and FuXi-ENS. Year-to-year analysis reveals substantial variability in forecast skill. SEAS5 performed exceptionally well in 2018 and 2019, when nearly all ensemble members aligned closely with observations. In these years, SEAS5 served as a reliable reference for identifying representative members from the other two forecasting systems. However, when SEAS5 skill declined, as in 2020, the value of this selection strategy diminished, underscoring the limitation of relying on a single benchmark. While SEAS5 remains the best available reference system, its occasional lack of skill suggests that ensemble selection methods should be made more adaptive, potentially by incorporating additional predictors such as large-scale circulation patterns (e.g., 500-hPa geopotential height) or combining multiple global systems as references to ensure more robust member selection. This will be the focus of our future work. Dynamical downscaling enhanced forecast performance by adding spatial detail and better capturing topographically driven variability. Downscaled FuXi-ENS successfully reproduced colder conditions over mountains and warmer conditions over plains, reducing the systematic cold bias of the global forecast and providing valuable regional insights. CFSv2 also benefited from downscaling, with substantial improvements in spatial correlation with observations at station locations. The ability of downscaling to add critical spatial detail is valuable for regional applications, especially in South Korea where complex topography strongly modulates local climate conditions. While its effectiveness ultimately depends on the quality of the input forecasts used as lateral boundary conditions, dynamical downscaling remains an indispensable step for translating global forecasts into actionable local information. The strong performance of FuXi-ENS further underscores the promise of ML–based forecasting. This study is the first to evaluate FuXi-ENS for extended-range forecasts beyond its originally intended 15-day horizon, and the results clearly demonstrate skillful performance up to one month. When combined with WRF, it gained additional spatial realism, producing a hybrid system with clear potential for operational use. This synergy highlights a promising path forward by coupling global ML-based forecasts with regional dynamical downscaling to leverage the strengths of both approaches. One methodological limitation is that, while the observed daily Tmean was computed from 24 hourly values, the model-based daily Tmean was derived from only four values per day. While this discrepancy has a negligible effect at monthly timescales, future improvements could be achieved if forecast models provided 1-hourly outputs or daily mean values directly, ensuring greater consistency with observational definitions. Addressing this issue would be particularly important for extending the framework to sub-daily variables or extremes. Although this study focused exclusively on daily Tmean to establish a baseline evaluation framework, the approach is readily extensible to other variables. As aforementioned, for example, large-scale circulation fields such as 500 hPa geopotential height, which are strongly linked to both temperature and precipitation, could be incorporated in member selection and evaluation. This would allow a more comprehensive diagnosis of forecast reliability and provide pathways to expand the hybrid approach beyond temperature prediction alone. In summary, this study demonstrates the feasibility and value of integrating ML–based global forecasts with regional dynamical downscaling to improve one-month temperature predictions over South Korea. SEAS5 proved useful as a benchmark for ensemble selection, though its limitations highlight the need for adaptive or multi-system reference strategies. Dynamical downscaling consistently improved regional representation, while FuXi-ENS emerged as a particularly promising driver of hybrid systems. Together, these results illustrate a pathway toward more accurate, actionable, and operationally viable extended-range forecasts, providing a foundation for future research and practical implementation in climate-sensitive sectors. Declarations Competing interests The authors declare no competing interests. Funding This study was supported by Korea Environment Industry & Technology Institute (KEITI) through Water Management Program for Drought, funded by Korea Ministry of Environment (MOE) (2480000175). Author Contribution Eun-Soon Im conceived and designed the study. Subin Ha collected the data, conducted the analysis, and wrote the original manuscript draft. Xiaohui Zhong, Lei Chen, and Hao Li provided the FuXi ensemble forecast data. Eun-Soon Im, Xiaohui Zhong, Jina Hur and Hyun-Han Kwon reviewed and edited the manuscript. Data Availability The observed data can be accessed at the data portal of the Korean Meteorological Administration ( [https://data.kma.go.kr/](https:/data.kma.go.kr) ). SEAS5 and CFSv2 forecasts are available at the ECMWF Copernicus Climate Data Store (CDS, [https://cds.climate.copernicus.eu/datasets/seasonal-original-single-levels?tab=overview](https:/cds.climate.copernicus.eu/datasets/seasonal-original-single-levels?tab=overview) ) and the NOAA NCEP CFS-model data archive ( [https://www.ncei.noaa.gov/data/climate-forecast-system/access/operational-9-month-forecast/](https:/www.ncei.noaa.gov/data/climate-forecast-system/access/operational-9-month-forecast) ), respectively. FuXi-ENS ensemble forecast data are available from the corresponding author upon reasonable request. References Hudson, D., Alves, O., Hendon, H. H. & Marshall, A. G. Bridging the gap between weather and seasonal forecasting: intraseasonal forecasting for Australia. Quart. J. Royal Meteoro Soc. 137 , 673–689 (2011). Robertson, A. W., Kumar, A., Peña, M. & Vitart, F. Improving and Promoting Subseasonal to Seasonal Prediction. Bull. Am. Meteorol. Soc. 96 , ES49–ES53 (2015). Vitart, F. et al. The Subseasonal to Seasonal (S2S) Prediction Project Database. Bull. Am. Meteorol. Soc. 98 , 163–173 (2017). Vitart, F. & Robertson, A. W. The sub-seasonal to seasonal prediction project (S2S) and the prediction of extreme events. npj Clim. Atmos. Sci. 1 , 3 (2018). Portele, T. C. et al. Seasonal forecasts offer economic benefit for hydrological decision making in semi-arid regions. Sci. Rep. 11 , 10581 (2021). White, C. J. et al. Advances in the Application and Utility of Subseasonal-to-Seasonal Predictions. Bull. Am. Meteorol. Soc. 103 , E1448–E1472 (2022). White, C. J. et al. Potential applications of subseasonal-to‐seasonal (S2S) predictions. Meteorol. Appl. 24 , 315–325 (2017). Chevuturi, A. et al. Forecast skill of the Indian monsoon and its onset in the ECMWF seasonal forecasting system 5 (SEAS5). Clim. Dyn. 56 , 2941–2957 (2021). Crespi, A., Petitta, M., Marson, P., Viel, C. & Grigis, L. Verification and Bias Adjustment of ECMWF SEAS5 Seasonal Forecasts over Europe for Climate Service Applications. Climate 9 , 181 (2021). De Andrade, F. M. et al. Subseasonal Precipitation Prediction for Africa: Forecast Evaluation and Sources of Predictability. Weather Forecast. 36 , 265–284 (2021). DeFlorio, M. J. et al. Experimental Subseasonal-to‐Seasonal (S2S) Forecasting of Atmospheric Rivers Over the Western United States. JGR Atmos. 124 , 11242–11265 (2019). Ehsan, M. A. et al. Seasonal predictability of Ethiopian Kiremt rainfall and forecast skill of ECMWF’s SEAS5 model. Clim. Dyn. 57 , 3075–3091 (2021). Ferreira, G. W. S., Reboita, M. S. & Drumond, A. Evaluation of ECMWF-SEAS5 Seasonal Temperature and Precipitation Predictions over South America. Climate 10 , 128 (2022). Siegmund, J., Bliefernicht, J., Laux, P. & Kunstmann, H. Toward a seasonal precipitation prediction system for West Africa: Performance of CFSv2 and high-resolution dynamical downscaling. JGR Atmos. 120 , 7316–7339 (2015). Tian, D., Wood, E. F. & Yuan, X. CFSv2-based sub-seasonal precipitation and temperature forecast skill over the contiguous United States. Hydrol. Earth Syst. Sci. 21 , 1477–1490 (2017). Wang, Q. J. et al. An evaluation of ECMWF SEAS5 seasonal climate forecasts for Australia using a new forecast calibration algorithm. Environ. Model. Softw. 122 , 104550 (2019). Weber, N. J. & Mass, C. F. Evaluating CFSv2 Subseasonal Forecast Skill with an Emphasis on Tropical Convection. Mon. Weather Rev. 145 , 3795–3815 (2017). Mandal, R. et al. Real time extended range prediction of heat waves over India. Sci. Rep. 9 , 9008 (2019). Kueh, M. T. & Lin, C. Y. The 2018 summer heatwaves over northwestern Europe and its extended-range prediction. Sci. Rep. 10 , 19283 (2020). Kim, O. & Lee, W. S. Enhancing seasonal predictability of the East Asian summer monsoon via optimized multi-model ensembles. Sci. Rep. 15 , 24912 (2025). Ha, S., Wang, Q. J., Hur, J. & Im, E. S. Evaluating interannual variability in 1-month CFSv2 forecasts with different lead times at monthly and daily scales in South Korea. Environ. Res. Commun. 7 , 041013 (2025). Im, E. S. et al. An Evaluation of Temperature-Based Agricultural Indices Over Korea From the High-Resolution WRF Simulation. Front. Earth Sci. 9 , 656787 (2021). Lee, Y., Pianosi, F., Peñuela, A. & Rico-Ramirez, M. A. Skill of seasonal flow forecasts at catchment scale: an assessment across South Korea. Hydrol. Earth Syst. Sci. 28 , 3261–3279 (2024). Lee, Y., Peñuela, A., Pianosi, F. & Rico-Ramirez, M. A. Catchment‐scale skill assessment of seasonal precipitation forecasts across South Korea. Intl J. Climatology . 43 , 5092–5111 (2023). Oh, J., Oh, J. & Huh, M. A. Case Study of Bias Correction and the Dynamical Downscaling of CFSv2 S2S Forecasts Using a WRF Model: Heatwave in 2018 over South Korea. Atmosphere 13 , 2086 (2022). Chen, L. et al. A machine learning model that outperforms conventional global subseasonal forecast models. Nat. Commun. 15 , 6425 (2024). Chen, L. et al. FuXi: a cascade machine learning forecasting system for 15-day global weather forecast. npj Clim. Atmos. Sci. 6 , 190 (2023). Kim, D. H., Moon, I. J. & Joh, M. Recent Advances in AI-based Atmospheric General Circulation Models. Atmosphere 35 , 263–276 (2025). Li, Y., Lu, G., Wu, Z., He, H. & He, J. High-Resolution Dynamical Downscaling of Seasonal Precipitation Forecasts for the Hanjiang Basin in China Using the Weather Research and Forecasting Model. J. Appl. Meteorol. Climatology . 56 , 1515–1536 (2017). Zhong, X. et al. FuXi-ENS: A machine learning model for medium-range ensemble weather forecasting. Preprint at. https://doi.org/10.48550/ARXIV.2405.05925 (2024). Hwang, J., Orenstein, P., Cohen, J., Pfeiffer, K. & Mackey, L. Improving Subseasonal Forecasting in the Western U.S. with Machine Learning. in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining 2325–2335 (ACM, Anchorage AK USA, (2019). 10.1145/3292500.3330674 Kenne, A. D., Toure, M., Logamou Seknewna, L. & Ketsemen, H. L. Subseasonal Prediction of Summer Temperature in West Africa Using Artificial Intelligence: A Case Study of Senegal. Int. J. Intell. Syst. 2024 , 1–19 (2024). Wang, C. et al. Improving the Accuracy of Subseasonal Forecasting of China Precipitation With a Machine Learning Approach. Front. Earth Sci. 9 , 659310 (2021). Zhang, L. et al. Improving Subseasonal-to-Seasonal forecasts in predicting the occurrence of extreme precipitation events over the contiguous U.S. using machine learning models. Atmos. Res. 281 , 106502 (2023). Pinheiro, E. & Ouarda, T. B. M. J. Short-lead seasonal precipitation forecast in northeastern Brazil using an ensemble of artificial neural networks. Sci. Rep. 13 , 20429 (2023). Kim, Y. I., Kim, D. & Lee, S. O. Prediction of Temperature and Heat Wave Occurrence for Summer Season Using Machine Learning. J. Korean Soc. Disaster Secur. 13 , 27–38 (2020). Oh, S. G. et al. Deep learning model for heavy rainfall nowcasting in South Korea. Weather Clim. Extremes . 44 , 100652 (2024). Oh, S. G. et al. Evaluation of Deep-Learning-Based Very Short-Term Rainfall Forecasts in South Korea. Asia-Pac J. Atmos. Sci. 59 , 239–255 (2023). Copernicus Climate Change Service. Seasonal forecast daily and subdaily data on single levels. ECMWF (2018). https://doi.org/10.24381/CDS.181D637E Saha, S. et al. The NCEP Climate Forecast System Version 2. J. Clim. 27 , 2185–2208 (2014). Gubler, S. et al. Assessment of ECMWF SEAS5 Seasonal Forecast Performance over South America. Weather Forecast. 35 , 561–584 (2019). Qiu, L., Im, E. S., Hur, J. & Shim, K. M. Added value of very high resolution climate simulations over South Korea using WRF modeling system. Clim. Dyn. 54 , 173–189 (2020). Hong, S. Y., Dudhia, J. & Chen, S. H. A Revised Approach to Ice Microphysical Processes for the Bulk Parameterization of Clouds and Precipitation. Mon Wea Rev. 132 , 103–120 (2004). Iacono, M. J. et al. Radiative forcing by long-lived greenhouse gases: Calculations with the AER radiative transfer models. J. Geophys. Res. 113 , 2008JD009944 (2008). Jiménez, P. A. et al. A Revised Scheme for the WRF Surface Layer Formulation. Mon. Weather Rev. 140 , 898–918 (2012). Chen, F. & Dudhia, J. Coupling an Advanced Land Surface–Hydrology Model with the Penn State–NCAR MM5 Modeling System. Part I: Model Implementation and Sensitivity. Mon Wea Rev. 129 , 569–585 (2001). Hong, S. Y., Noh, Y. & Dudhia, J. A. New Vertical Diffusion Package with an Explicit Treatment of Entrainment Processes. Mon. Weather Rev. 134 , 2318–2341 (2006). Kain, J. S. The Kain–Fritsch Convective Parameterization: An Update. J. Appl. Meteor. 43 , 170–181 (2004). Ma, L. M. & Tan, Z. M. Improving the behavior of the cumulus parameterization for tropical cyclone prediction: Convection trigger. Atmos. Res. 92 , 190–211 (2009). Additional Declarations No competing interests reported. Supplementary Files SRSupplementary.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 29 Jan, 2026 Reviews received at journal 12 Dec, 2025 Reviews received at journal 09 Dec, 2025 Reviewers agreed at journal 11 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers agreed at journal 16 Oct, 2025 Reviewers agreed at journal 25 Sep, 2025 Reviewers invited by journal 24 Sep, 2025 Editor invited by journal 24 Sep, 2025 Editor assigned by journal 23 Sep, 2025 Submission checks completed at journal 19 Sep, 2025 First submitted to journal 17 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7639295","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":524739171,"identity":"36c2f52b-ea5e-4996-91fd-69db234336b7","order_by":0,"name":"Subin Ha","email":"","orcid":"","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Subin","middleName":"","lastName":"Ha","suffix":""},{"id":524739172,"identity":"3d28c4ef-076d-451f-881f-0bfdcedd5d34","order_by":1,"name":"Xiaohui Zhong","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Xiaohui","middleName":"","lastName":"Zhong","suffix":""},{"id":524739173,"identity":"dbbdf8be-b622-467f-8b25-77af267db64c","order_by":2,"name":"Lei Chen","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Chen","suffix":""},{"id":524739174,"identity":"43b49682-4b77-4b96-a3aa-13c7b60f4afc","order_by":3,"name":"Hao Li","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Hao","middleName":"","lastName":"Li","suffix":""},{"id":524739175,"identity":"c8261a13-d046-4952-aa01-4a736e845ad1","order_by":4,"name":"Jina Hur","email":"","orcid":"","institution":"National Institute of Agricultural Sciences, Rural Development Administration","correspondingAuthor":false,"prefix":"","firstName":"Jina","middleName":"","lastName":"Hur","suffix":""},{"id":524739176,"identity":"057777b8-d10e-443f-9099-e3b81c9c115c","order_by":5,"name":"Hyun-Han Kwon","email":"","orcid":"","institution":"Sejong University","correspondingAuthor":false,"prefix":"","firstName":"Hyun-Han","middleName":"","lastName":"Kwon","suffix":""},{"id":524739177,"identity":"ffe884dd-a29c-49b0-ba96-7da2ba8753b4","order_by":6,"name":"Eun-Soon Im","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAl0lEQVRIiWNgGAWjYFACxgaGDyBago0ELYwzgBpI0cLAwMxDkhb+9ubWzbZtdXUM0m0JxGmROHOw7XZu22EJBpljB4jTYiCRCNJyAOiw9AYitcg/bLtt2VZHihYJxrbbjG3MQC1pRDpM4kxi282ec4cl2yTSEojTwt9+/NmNH2V1/PwSaQbEaYEDUiJyFIyCUTAKRgEhAABRzygGLsK0KAAAAABJRU5ErkJggg==","orcid":"","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Eun-Soon","middleName":"","lastName":"Im","suffix":""}],"badges":[],"createdAt":"2025-09-17 10:53:30","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7639295/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7639295/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92931715,"identity":"08f2544f-5aee-40d2-8943-8e76e6ff5536","added_by":"auto","created_at":"2025-10-07 09:13:49","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3501786,"visible":true,"origin":"","legend":"","description":"","filename":"SRMain.docx","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/649c9ed34ebfdf889f321d29.docx"},{"id":92930734,"identity":"72ee8a55-1733-4edc-aa55-d92b70f29e29","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8523,"visible":true,"origin":"","legend":"","description":"","filename":"13be099fbff94f4c9b6c0c8719a53fa5.json","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/b913044ccf671469f8a60534.json"},{"id":92931714,"identity":"4aaa2436-4070-4dac-b73b-b3c552dda34f","added_by":"auto","created_at":"2025-10-07 09:13:49","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1332830,"visible":true,"origin":"","legend":"","description":"","filename":"SRSupplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/affd3fec08bb89b0a762cefa.docx"},{"id":92932147,"identity":"ba4a4362-f658-4c59-a971-be2a861083c9","added_by":"auto","created_at":"2025-10-07 09:21:49","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":104208,"visible":true,"origin":"","legend":"","description":"","filename":"13be099fbff94f4c9b6c0c8719a53fa51enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/ae27c389f7bccfe2aea39860.xml"},{"id":92931712,"identity":"d9512236-0383-4b7d-80b4-0217f1263e24","added_by":"auto","created_at":"2025-10-07 09:13:49","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":579342,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/1e0214d3c914c93a666e44ba.jpeg"},{"id":92930751,"identity":"824728af-1cd3-4e70-8dba-738c52a72f66","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1474910,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/58acdf3fb41ad18f690715cb.jpeg"},{"id":92930740,"identity":"282fbb9a-68e4-4e61-9e78-aeceb3a5ee02","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137125,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/33da98af39d1f85f3c775ce0.png"},{"id":92930757,"identity":"0d205a48-b1ae-4ac3-9ded-ff1d5fdbb3e4","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40507,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/a6e7e64283d64e779fd7732b.jpeg"},{"id":92930756,"identity":"db3b27aa-74ba-4239-8db1-a5033f742139","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":80027,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/42703ffd345709297637814c.png"},{"id":92930752,"identity":"72ba6742-cf39-470c-a025-f7e6410f0467","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":321622,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/ad1af70adbf4d55f07986610.png"},{"id":92930753,"identity":"897f2c25-e4ed-4305-95bf-9eaeb441b007","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":34631,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/b53dc5f4956ff769668a19d2.png"},{"id":92930743,"identity":"ba2c9bf5-bca1-48d8-800a-318f9918b019","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13671,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/b46488f58b4e927933d97003.png"},{"id":92930745,"identity":"85c4dfd9-6baa-4347-84f4-faf920d98aa7","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"xml","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":103073,"visible":true,"origin":"","legend":"","description":"","filename":"13be099fbff94f4c9b6c0c8719a53fa51structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/e37afdc1443013b75a7a88ca.xml"},{"id":92930754,"identity":"e9218e46-8675-41f6-a8ab-23e7bc290d3a","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"html","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":113920,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/eb8c16971f8301643c2271b8.html"},{"id":92930733,"identity":"29bfd47a-4121-4ad4-ae41-1b484628d98e","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":326059,"visible":true,"origin":"","legend":"\u003cp\u003eBoxplots of (a) correlation coefficients and (b) RMSEs for SEAS5 (51 members), CFSv2 (40 members), and FuXi-ENS (40 members) against OBS for daily Tmean. (Some CFSv2 forecast data are missing from the data archive, possibly due to upload issues, with 5, 11, 28, and 1 out of 40 members missing in 2021, 2022, 2023, and 2024, respectively.)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/25b3eae2bba23724ff0e77df.png"},{"id":92930735,"identity":"83a5075c-3014-4504-8035-02c261737071","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":488728,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmaps ofdaily Tmean correlation coefficients of SEAS5 (51 members), CFSv2 (40 members), FuXi-ENS (40 members) with OBS (Low correlation coefficients between -0.2 and 0.2 are colored in white. The green boxes indicate the CFSv2 and FuXi-ENS forecast members selected based on the comparison with SEAS5-ENS.)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/0d047d4038956b08058ea169.png"},{"id":92930736,"identity":"8cab4e1f-668a-4c77-9241-5f0f1d05463b","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":321622,"visible":true,"origin":"","legend":"\u003cp\u003eWeekly Tmean spatial maps of OBS, SEAS5-ENS, CFS-OG-ENS4, CFS-DS-ENS4, FuXi-OG-ENS4, and FuXi-DS-ENS4 for July 2018 and 2020 (Week 1/2/3/4: July 4-10, 11-17, 18-24, 25-31; Spatial mean values are shown in the top right corners.)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/0de405508b445e52d17aa447.png"},{"id":92930737,"identity":"bd88cbd4-6392-422d-82aa-d8ffd0577e51","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":139279,"visible":true,"origin":"","legend":"\u003cp\u003eTemporal correlation of daily Tmean (x-axis) versus spatial correlation of monthly Tmean (y-axis) for each ensemble member of CFSv2 and FuXi-ENS. (Different marker shapes denote years (2018, 2019, 2020); green markers indicate CFSv2 and red markers indicate FuXi-ENS; empty markers represent original forecasts (OG) and filled markers represent dynamically downscaled forecasts (DS). The light blue box highlights the region where both temporal and spatial correlations exceed 0.5.)\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/dbc8641735decbddd29bb6b3.png"},{"id":92933026,"identity":"cbc9d31c-a8a2-4ac0-abf2-979473117b7f","added_by":"auto","created_at":"2025-10-07 09:29:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1969275,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/cdc76357-cb97-4568-8e40-a649554ed32b.pdf"},{"id":92930739,"identity":"6ddfb7f0-d2b3-4a85-850a-1f7269d176df","added_by":"auto","created_at":"2025-10-07 09:05:49","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1332830,"visible":true,"origin":"","legend":"","description":"","filename":"SRSupplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-7639295/v1/4ffb04bb3a9d98a041312250.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Enhancing 1-month temperature predictions in South Korea through dynamical downscaling of machine learning global ensemble forecasts","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSubseasonal-to-seasonal (S2S) forecasting plays an increasingly important role in bridging the critical gap between short-term weather forecasts and long-term seasonal outlooks\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Accurate forecasts at this timescale are essential for effective resource planning and risk management in various sectors such as agriculture, water resource management, energy, public health, and disaster preparedness, where decisions often depend on climate conditions weeks to months in advance\u003csup\u003e\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. In response to this need, global seasonal forecasting systems have been developed and extensively assessed across regions and variables, ranging from near-surface meteorological variables like temperature and precipitation to large-scale climate phenomena such as El Ni\u0026ntilde;o and seasonal monsoons\u003csup\u003e\u003cspan additionalcitationids=\"CR9 CR10 CR11 CR12 CR13 CR14 CR15 CR16 CR17 CR18 CR19\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. For South Korea, several studies have assessed the performance and utility of various seasonal forecasts\u003csup\u003e\u003cspan additionalcitationids=\"CR22 CR23 CR24\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. Using the Climate Forecast System version 2 (CFSv2) operational forecasts, Ha et al.\u003csup\u003e21\u003c/sup\u003e found that forecasts with shorter lead times are more likely to perform better in capturing interannual temperature variability of South Korea. Still, the relatively coarse spatial resolution of global forecasts remains a major obstacle, limiting their ability to represent localized variability and hindering their direct application for end users. To address these shortcomings, dynamical downscaling has been utilized by a few studies to enhance regional details through high-resolution simulations. Im et al.\u003csup\u003e22\u003c/sup\u003e and Oh et al.\u003csup\u003e25\u003c/sup\u003e showed that applying dynamical downscaling to CFSv2 forecasts significantly improves their temperature prediction for the country, enabling better use of forecasts for downstream applications such as calculating temperature-based agricultural indices and predicting heatwaves. These improvements highlight the added value of dynamical downscaling of global forecasts.\u003c/p\u003e\u003cp\u003eWhile physical models have been the mainstay of global forecasting, machine learning (ML) has rapidly led to significant advances in forecasting across different timescales, including the S2S range. Recent studies have demonstrated that ML-based forecasts can match or surpass the performance of traditional physical models. In particular, a number of studies have demonstrated that the FuXi-ENS model outperforms the European Centre for Medium-Range Weather Forecasts (ECMWF)\u0026rsquo;s state-of-the-art system\u003csup\u003e\u003cspan additionalcitationids=\"CR27 CR28 CR29\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. Other ML approaches have also yielded improved temperature and precipitation forecasts in various regions, including China, Brazil, West Africa, and the U.S., consistently outperforming traditional physics-based models for subseasonal timescales\u003csup\u003e\u003cspan additionalcitationids=\"CR32 CR33 CR34\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. Despite these promising advances, applications of ML-based forecasting in South Korea remain scarce, and existing work is mostly limited to nowcasting or short-term prediction\u003csup\u003e\u003cspan additionalcitationids=\"CR37\" citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. As ML forecasting systems continue to advance, there is a pressing need to evaluate their performance in South Korea and explore ways to enhance their utility by integrating techniques such as dynamical downscaling in order to bridge the scaling gap between prediction and actionable information.\u003c/p\u003e\u003cp\u003eIn this regard, this study evaluates the potential of combining ML-based global forecasts with dynamical downscaling to improve 1-month temperature prediction over South Korea, with a focus on the month of July from 2018 to 2024. Specifically, we assess and compare the performance of three global forecasts: the Seasonal Forecasting System version 5 (SEAS5) from ECMWF, the CFSv2 from the National Centers for Environmental Prediction (NCEP), and the ML-based FuXi-ENS model developed by Fudan University. The forecast skill of each system is first evaluated against observed daily temperature data in South Korea. To examine the added value of high-resolution simulations, dynamical downscaling is then applied to selected ensemble members from CFSv2 and FuXi. The downscaled outputs are subsequently evaluated against observations. As far as our knowledge extends, this is the first attempt to apply dynamical downscaling to an ML-based global forecast system to obtain fine-scale forecasting information with a 1-month lead time targeted at South Korea. Moreover, by focusing on daily temperature forecasts rather than monthly averages, this study aligns more closely with the needs of end users, who often require high-frequency, locally actionable information. This comparative assessment will provide valuable insights into the feasibility and effectiveness of integrating readily available global forecasts and a customized dynamical downscaling system for the target region for improving subseasonal forecasting, ultimately contributing to more informed and timely decision-making in climate-sensitive sectors.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStudy domain and observed data (OBS)\u003c/h2\u003e\u003cp\u003eThis study focuses on one-month temperature prediction on a daily timescale for July during 2018\u0026ndash;2024 in South Korea. OBS for this period were obtained from the Korean Meteorological Administration. Each station measures 2-m air temperature at hourly intervals, and the daily mean temperature (Tmean) was computed as the average of the 24 values. Tmean records from 506 observation stations were used in this study, including 88 Automated Synoptic Observing System (ASOS) stations and 418 Automatic Weather System (AWS) stations. The spatial distribution of these stations is shown in Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e online.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eGlobal seasonal forecasts\u003c/h3\u003e\n\u003cp\u003eThree global seasonal forecast models are employed in this study: SEAS5, CFSv2, and FuXi-ENS. SEAS5\u003csup\u003e39\u003c/sup\u003e is the latest seasonal forecasting system developed by ECMWF, consisting of a 51-member ensemble. Operational since November 2017, SEAS5 is initialized on the 1st of each month and provides forecasts extending up to 7 months. Surface-level data are available at 6-hour intervals, while pressure-level data are available at 12-hour intervals. Due to the coarse temporal resolution of upper-level fields, SEAS5 cannot be directly used as input for dynamical downscaling. Instead, 6-hourly 2-meter temperature data at a horizontal resolution of 1\u0026deg; are employed as a benchmark for selecting ensemble members from CFSv2 and FuXi-ENS for dynamical downscaling. Daily Tmean from SEAS5 is derived from the four available 6-hourly values each day.\u003c/p\u003e\u003cp\u003eCFSv2\u003csup\u003e40\u003c/sup\u003e operational forecasts by NCEP have been available since April 2011 and are publicly available through the NCEP data archive. The forecasting system is initialized four times daily (00, 06, 12, and 18 UTC) and produces deterministic forecasts up to nine months at 6-hour intervals. Surface-level forecast data are available at approximately 100 km resolution (Gaussian T126), while pressure-level data are provided at 0.5\u0026deg; resolution. For this study, 40 ensemble members initialized between June 21st UTC00 and June 30th UTC18 for each year are used to generate July forecasts. Daily Tmean is obtained by averaging the four 6-hourly forecasts available per day.\u003c/p\u003e\u003cp\u003eFuXi-ENS\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e is a machine learning-based seasonal forecasting system developed using ERA5 as training data, with a spatial resolution of 0.25\u0026deg;. Trained on ERA5 data through 2017, FuXi-ENS produces forecasts available from 2018 onward. While originally designed for 0\u0026ndash;15 day forecasts, here FuXi-ENS is extended to 33 days to assess its one-month prediction skill. Each year, 40 perturbed ensemble members were initialized at 00 UTC on June 29th, with each forecast providing 6-hourly data at both surface and pressure levels through the end of July. Like the other two aforementioned models, daily Tmean from FuXi-ENS is constructed as the arithmetic mean of its four 6-hourly outputs per day.\u003c/p\u003e\n\u003ch3\u003eStudy period and ensemble member selection for dynamical downscaling\u003c/h3\u003e\n\u003cp\u003eBoth CFSv2 and FuXi-ENS provide 40 ensemble members, respectively, for July forecasts each year. Due to computational constraints, it is not feasible to dynamically downscale all ensemble members over the entire study period. Therefore, 4 representative members (10% of the ensemble) are selected from each forecast system per year for dynamical downscaling.\u003c/p\u003e\u003cp\u003eFor historical periods, observed temperature records are available, so ensemble members can be selected based on their agreement with OBS before downscaling. However, the ultimate goal of this research is to develop an operational system for producing practical 1-month forecasts for South Korea, applicable to future periods when OBS are not yet available. Thus, for a generalizable and operationally feasible member selection strategy, a proxy reference is required. SEAS5 is used as this benchmark based on its generally superior forecasting skills, as demonstrated in previous studies\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the correlation coefficients and root-mean-square errors (RMSEs) of daily Tmean between OBS and forecasts in South Korea from each global model for July in each year. In most years, SEAS5 outperforms CFSv2 and FuXi-ENS, showing higher correlation coefficients and smaller RMSEs against OBS. Furthermore, despite having more ensemble members (51) than CFSv2 and FuXi-ENS (40), SEAS5 generally exhibits smaller inter-member variability, as indicated by narrower boxplots for both correlation and RMSE.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWhile these results generally support the use of SEAS5 as a proxy for observed conditions in member selection since 2018, when FuXi-ENS forecasts became available, the dynamical downscaling of CFSv2 and FuXi-ENS for comparative assessment is conducted for the years 2018, 2019, and 2020, when there is no missing data for CFSv2. Additionally, these three consecutive years represent a distinct range of July temperature conditions: in the recent 20 years (mean climatology: 24.6\u0026deg;C), 2018 experienced the hottest July on record (mean temperature: 26.6\u0026deg;C), 2020 the coolest (mean temperature: 22.5\u0026deg;C), and 2019 conditions close to the climatological average (mean temperature: 24.7\u0026deg;C). Model performance also varied markedly among these years. In 2018, the lower quartile of SEAS5 members\u0026rsquo; correlation coefficients exceeded 0.8, indicating consistently strong agreement with OBS across nearly the entire ensemble members; FuXi-ENS members also showed high correlations. In contrast, in 2020, even the upper quartile of SEAS5 correlations was only\u0026thinsp;~\u0026thinsp;0.25, highlighting particularly weak forecast performance; CFSv2 and FuXi-ENS exhibited similarly low correlations. Interestingly, RMSEs in 2020 were small for all three models, with narrow inter-member ranges. This indicates that while the models fail to capture daily fluctuations, their absolute errors are relatively small. The year 2019 displayed intermediate characteristics. These three years, therefore, represent cases with distinct observed temperature conditions and differing forecast skill across the three systems, warranting detailed analysis in the following sections.\u003c/p\u003e\u003cp\u003eThe selection process proceeds as follows: For each year, daily Tmean values of the 51 SEAS5 members are averaged to form the ensemble mean, hereafter referred to as SEAS5-ENS. Then, the daily Tmean of each CFSv2 and FuXi-ENS ensemble member is compared to SEAS5-ENS by calculating the Pearson correlation coefficient for July. The 4 members from each model with the highest correlations to SEAS5-ENS are selected for dynamical downscaling.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel configuration for dynamical downscaling\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe Weather and Research Forecasting (WRF) model version 4.5 is employed for dynamical downscaling of CFSv2 and FuXi-ENS forecasts. A one-way nested domain setup is used, consisting of a mother domain at 20 km resolution and a nested domain at 5 km resolution, both defined using a Lambert conformal projection centered at 38\u0026deg;N and 127.5\u0026deg;E (see Supplementary Fig. S2 online). The physical parameterization schemes follow the optimal configuration identified by Qiu et al.\u003csup\u003e42\u003c/sup\u003e through sensitivity testing. Specifically, the model employs the WSM3 microphysics scheme\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e, the RRTMG longwave and shortwave radiation schemes\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e, the Revised MM5 surface layer scheme\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e, the Noah land surface model\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e, the Yonsei University planetary boundary layer scheme\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e, and the Kain\u0026ndash;Fritsch cumulus convection scheme\u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e with trigger function of Ma and Tan\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eSuccessful initialization of WRF requires a minimum set of meteorological and land-surface fields. While CFSv2 forecasts provide all necessary variables, FuXi-ENS lacks several key fields, including skin temperature, surface pressure, soil moisture, and soil temperature. To address this gap, ERA5 reanalysis data are incorporated. Rather than prescribing daily variability, we construct July climatologies by averaging ERA5 data from 2015 to 2024 at four synoptic times (00, 06, 12, and 18 UTC) to align with the forecast intervals of CFSv2 and FuXi-ENS. This approach assumes that the predictability of FuXi-ENS downscaling primarily derives from the large-scale patterns, while acknowledging that day-to-day variations in surface variables may not be fully captured.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003eSkill evaluation of SEAS5, CFSv2, and FuXi-ENS\u003c/h2\u003e\u003cp\u003eThe correlation between Tmean from each ensemble member of SEAS5, CFSv2, and FuXi-ENS with OBS for the study period is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e as heatmaps. Consistent with Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the superior performance of SEAS5 is evident, with a dominance of strong red colors indicating high positive correlations. While most members of CFSv2 and FuXi-ENS also exhibit positive correlations with OBS, their intensities are generally weaker, shown by lighter red shades. Notably, a considerable number of members, particularly from CFSv2, are shown in white or near-white, indicating low correlation, and some even appear in blue, highlighting negative correlations with OBS.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWhen comparing CFSv2 and FuXi-ENS, FuXi-ENS appears to slightly outperform CFSv2 overall. However, it's important to note a key distinction in initialization: CFSv2 members are initialized every 6 hours, resulting in different lead times ranging from 240 hours to 6 hours, whereas FuXi-ENS ensemble members are all initialized at 00 UTC on June 29, corresponding to a fixed lead time of 48 hours, with perturbations used to generate ensemble spread. In Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, for each year, CFSv2 members are ordered from top to bottom based on lead time, with the longest lead time (ddhh: 2100) on the top and the shortest (ddhh: 3018) on the bottom. It is clearly seen that members with shorter lead times tend to show stronger positive correlations with OBS. This pattern aligns with findings from Ha et al.\u003csup\u003e21\u003c/sup\u003e, which showed that shorter lead times generally yield better temperature forecast skill.\u003c/p\u003e\u003cp\u003eOn a year-by-year basis, 2018 is especially notable, as all 51 SEAS5 members exhibit very strong positive correlations with OBS. Similarly, in 2019, most SEAS5 members are also strongly correlated. For these two years, FuXi-ENS ensemble members are predominantly positively correlated, and CFSv2 members with short lead times also show good correlation. In contrast, 2020 presents the weakest performance, with most SEAS5 members showing only weak positive correlations, and CFSv2 and FuXi-ENS members showing minimal skill compared to the previous two years.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eMember selection for dynamical downscaling\u003c/h2\u003e\u003cp\u003eFollowing the member selection procedure outlined in Section 2.3.1, SEAS5-ENS is used to compute daily Tmean correlation coefficients with each CFSv2 and FuXi-ENS member. The four members from each system with the highest correlation to SEAS5-ENS are selected for dynamical downscaling and are indicated by green boxes in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The details of selected members, including initialization times for CFSv2 and member numbers for FuXi-ENS, are provided in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e online. To evaluate the selection method, the actual correlation of the selected members with OBS is assessed, and each selected member\u0026rsquo;s ranking in terms of correlation with OBS among the 40 total members is also provided in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e online to offer insight into the relative skill of selected and non-selected members.\u003c/p\u003e\u003cp\u003eThe effectiveness of the selection method is found to be highly dependent on the predictive skill of SEAS5, which is used as the selection benchmark. In 2018, for example, the selected CFSv2 and FuXi-ENS members\u0026mdash;based on correlation with SEAS5-ENS\u0026mdash;also turn out to be the top 4 members most highly correlated with OBS. For FuXi-ENS, the ranking of selected members based on SEAS5-ENS matches exactly with their ranking against OBS (1st to 4th), demonstrating a perfect selection outcome. This success reflects the high skill of SEAS5 forecasts in 2018, where all 51 SEAS5 members showed strong positive correlations with OBS. Consequently, SEAS5 served as a reliable reference for guiding member selection. Similarly, in 2019, when SEAS5 forecasts were also skillful, FuXi-ENS member selection performed well, successfully identifying the top 4 OBS-correlated members. The CFSv2 selection in 2019 was also reasonable, capturing the members ranked 3rd, 5th, 6th, and 7th in correlation with OBS. However, in 2020, when SEAS5 members had the weakest correlations with OBS, the selected members from both CFSv2 and FuXi-ENS mostly ranked in the bottom 50% in terms of their actual correlation with OBS, highlighting the challenge of making accurate selections when the benchmark itself lacks predictive skill.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eAdded value of dynamically downscaled forecasts\u003c/h3\u003e\n\u003cp\u003eThe selected CFSv2 and FuXi-ENS ensemble members were dynamically downscaled using the WRF model configuration described in Section 2.3.2. The mean of the four selected original CFSv2 members is referred to as CFS-OG-ENS4 (OG: original), and the mean of their dynamically downscaled outputs is referred to as CFS-DS-ENS4 (DS: downscaled). Similarly, the FuXi-based original and downscaled ensemble means are denoted as FuXi-OG-ENS4 and FuXi-DS-ENS4, respectively. To evaluate the added value of dynamical downscaling, we first assess the model performance in simulating regionally distinct temperature patterns. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the spatial distribution of weekly Tmean for OBS, SEAS5-ENS, CFS-OG-ENS4, CFS-DS-ENS4, FuXi-OG-ENS4, and FuXi-DS-ENS4 for July in 2018 and 2020; results for July 2019 are provided in Supplementary Fig. S3 online.\u003c/p\u003e\u003cp\u003eIn July 2018, temperatures increased steadily from week 1 to week 4. In OBS, the spatial mean was around 22\u0026deg;C in week 1, rose sharply to almost 27\u0026deg;C in week 2, and remained high at 28.1\u0026deg;C and 28.6\u0026deg;C in weeks 3 and 4, respectively. This warming trend was generally reproduced by SEAS5, CFSv2, and FuXi-ENS, although all three underestimated absolute temperatures due to their coarse resolution. SEAS5 and CFSv2, in particular, simulated relatively uniform temperatures across the country, which primarily follow latitudinal gradients, and failed to capture topographically driven variations. In contrast, FuXi-ENS better differentiated regional temperature patterns and exhibited the smallest cold bias, likely due to its comparatively higher native resolution. Dynamical downscaling further improved FuXi-ENS\u0026rsquo;s spatial representation, capturing colder conditions over mountainous areas and warmer conditions over lowland plains more accurately. Although the spatial mean temperature decreased slightly after downscaling due to the better-resolved cool mountain areas, the enhanced spatial detail provides valuable regional-scale information beyond that available in the original forecasts. CFSv2 also shows enhanced spatial details after downscaling, but its large cold bias in the original forecasts persisted. In weeks 3 and 4, when OBS exceeded 28\u0026deg;C, CFS-DS-ENS4 remained more than 3.5\u0026deg;C cooler than OBS, whereas FuXi-DS-ENS4 was approximately 2\u0026deg;C cooler.\u003c/p\u003e\u003cp\u003eIn July 2020, observed conditions did not follow a steady warming trend. OBS recorded 22.7\u0026deg;C in week 1, followed by a drop of about 2\u0026deg;C in week 2, a return to week-1 levels in week 3, and only a slight increase (0.5\u0026deg;C) in week 4. This fluctuating pattern was not captured by any of the three original forecast systems, all of which predicted a monotonic week-to-week warming. This suggests that while forecast systems can predict steadily warming conditions (e.g., 2018), they may be less effective at capturing complex temporal fluctuations such as those observed in 2020. After downscaling, FuXi-DS-ENS4 showed a warm bias of only 0.8\u0026deg;C in week 1 but failed to capture the sharp week-2 cooling, increasing the warm bias to 4.6\u0026deg;C, and in weeks 3 and 4, FuXi-DS-ENS4 maintained a warm bias of ~\u0026thinsp;3\u0026deg;C. Similarly, while CFS-DS-ENS4 matched OBS well in week 1, it failed to capture subsequent fluctuations, maintaining a warm bias thereafter. Thus, while spatial details improve after dynamical downscaling, the models struggled to reproduce weekly temperature variability, reflecting that the primary source of predictive skill resides in the original large-scale forecasts.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn addition to the evaluation of spatial temperature patterns, we further assessed the forecasts in terms of their temporal consistency and spatial fidelity against observations. Specifically, for each of the four ensemble members that constitute ENS4, we computed the temporal correlation of daily Tmean throughout July with OBS and the spatial correlation of monthly Tmean with OBS. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents daily Tmean temporal correlation on the x-axis and monthly Tmean spatial correlation on the y-axis for each ensemble member of CFSv2 and FuXi-ENS.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAs already demonstrated in the preceding results, the year 2018 exhibited markedly higher forecast skill compared to 2020, a finding that is again confirmed in this figure. In 2018, both CFSv2 and FuXi-ENS members achieved substantially higher temporal correlations, clustering toward the far right side of the x-axis. Notably, CFSv2 benefitted substantially from dynamical downscaling. While temporal correlations remained largely unchanged before and after downscaling, spatial correlations increased from around 0.5 to 0.8. This improvement may be attributed to the refinement from the coarse 1\u0026deg; grid of the original forecasts to the 5 km grid of the downscaled products, which allows each observational station to be represented by values that more closely match local conditions. A similar benefit was found in 2020, with CFSv2 ensemble members improving approximately from 0.4 to 0.8 in spatial correlation after downscaling, while maintaining similar levels of temporal correlation.\u003c/p\u003e\u003cp\u003eAt first glance, FuXi-ENS appears to gain relatively little from dynamical downscaling, since the increases in station-based spatial correlation are modest. However, the correlation metrics in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e are calculated only at available station locations, whereas the spatial maps shown earlier (i.e., Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) clearly demonstrate that FuXi-ENS benefits significantly from dynamical downscaling in terms of improved representation of spatially distinct temperature features.\u003c/p\u003e\u003cp\u003eIt should also be noted that FuXi-ENS, which already possesses relatively high native resolution, exhibited much higher correlations even before downscaling, with its original forecasts already achieving levels comparable to the downscaled CFSv2 products. Nonetheless, after downscaling, FuXi-ENS still exhibited modest but consistent spatial gains, particularly in 2020 when the original forecasts performed less well, highlighting the added value of higher-resolution simulations. Although a slight decrease in temporal correlation was observed for some FuXi-ENS members, the magnitude of this reduction was rather small and does not materially affect the overall temporal fidelity of the downscaled products.\u003c/p\u003e\u003cp\u003eOverall, these results emphasize that while the general predictability of temperature variability originates from the global forecasts themselves, dynamical downscaling enhances the realism of regional-scale representation. The downscaled products inherit the temporal patterns of the original forecasts but resolve local heterogeneity more accurately, producing forecasts that are more realistic and locally relevant. This highlights the complementary role of dynamical downscaling in bridging the gap between global predictability and regional usability.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study evaluated the performance of 1-month temperature forecasts for July over South Korea using three global forecast systems, SEAS5, CFSv2, and FuXi-ENS, focusing on the years 2018, 2019, and 2020, which represent contrasting temperature patterns. Comparisons with observations from 506 ground-based stations show that SEAS5 consistently exhibits the highest correlation with daily Tmean, making it a practical benchmark for ensemble member selection and subsequent dynamical downscaling of CFSv2 and FuXi-ENS.\u003c/p\u003e\u003cp\u003eYear-to-year analysis reveals substantial variability in forecast skill. SEAS5 performed exceptionally well in 2018 and 2019, when nearly all ensemble members aligned closely with observations. In these years, SEAS5 served as a reliable reference for identifying representative members from the other two forecasting systems. However, when SEAS5 skill declined, as in 2020, the value of this selection strategy diminished, underscoring the limitation of relying on a single benchmark. While SEAS5 remains the best available reference system, its occasional lack of skill suggests that ensemble selection methods should be made more adaptive, potentially by incorporating additional predictors such as large-scale circulation patterns (e.g., 500-hPa geopotential height) or combining multiple global systems as references to ensure more robust member selection. This will be the focus of our future work.\u003c/p\u003e\u003cp\u003eDynamical downscaling enhanced forecast performance by adding spatial detail and better capturing topographically driven variability. Downscaled FuXi-ENS successfully reproduced colder conditions over mountains and warmer conditions over plains, reducing the systematic cold bias of the global forecast and providing valuable regional insights. CFSv2 also benefited from downscaling, with substantial improvements in spatial correlation with observations at station locations. The ability of downscaling to add critical spatial detail is valuable for regional applications, especially in South Korea where complex topography strongly modulates local climate conditions. While its effectiveness ultimately depends on the quality of the input forecasts used as lateral boundary conditions, dynamical downscaling remains an indispensable step for translating global forecasts into actionable local information.\u003c/p\u003e\u003cp\u003eThe strong performance of FuXi-ENS further underscores the promise of ML\u0026ndash;based forecasting. This study is the first to evaluate FuXi-ENS for extended-range forecasts beyond its originally intended 15-day horizon, and the results clearly demonstrate skillful performance up to one month. When combined with WRF, it gained additional spatial realism, producing a hybrid system with clear potential for operational use. This synergy highlights a promising path forward by coupling global ML-based forecasts with regional dynamical downscaling to leverage the strengths of both approaches.\u003c/p\u003e\u003cp\u003eOne methodological limitation is that, while the observed daily Tmean was computed from 24 hourly values, the model-based daily Tmean was derived from only four values per day. While this discrepancy has a negligible effect at monthly timescales, future improvements could be achieved if forecast models provided 1-hourly outputs or daily mean values directly, ensuring greater consistency with observational definitions. Addressing this issue would be particularly important for extending the framework to sub-daily variables or extremes.\u003c/p\u003e\u003cp\u003eAlthough this study focused exclusively on daily Tmean to establish a baseline evaluation framework, the approach is readily extensible to other variables. As aforementioned, for example, large-scale circulation fields such as 500 hPa geopotential height, which are strongly linked to both temperature and precipitation, could be incorporated in member selection and evaluation. This would allow a more comprehensive diagnosis of forecast reliability and provide pathways to expand the hybrid approach beyond temperature prediction alone.\u003c/p\u003e\u003cp\u003eIn summary, this study demonstrates the feasibility and value of integrating ML\u0026ndash;based global forecasts with regional dynamical downscaling to improve one-month temperature predictions over South Korea. SEAS5 proved useful as a benchmark for ensemble selection, though its limitations highlight the need for adaptive or multi-system reference strategies. Dynamical downscaling consistently improved regional representation, while FuXi-ENS emerged as a particularly promising driver of hybrid systems. Together, these results illustrate a pathway toward more accurate, actionable, and operationally viable extended-range forecasts, providing a foundation for future research and practical implementation in climate-sensitive sectors.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eCompeting interests\u003c/h2\u003e\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eThis study was supported by Korea Environment Industry \u0026amp; Technology Institute (KEITI) through Water Management Program for Drought, funded by Korea Ministry of Environment (MOE) (2480000175).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eEun-Soon Im conceived and designed the study. Subin Ha collected the data, conducted the analysis, and wrote the original manuscript draft. Xiaohui Zhong, Lei Chen, and Hao Li provided the FuXi ensemble forecast data. Eun-Soon Im, Xiaohui Zhong, Jina Hur and Hyun-Han Kwon reviewed and edited the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe observed data can be accessed at the data portal of the Korean Meteorological Administration ( [https://data.kma.go.kr/](https:/data.kma.go.kr) ). SEAS5 and CFSv2 forecasts are available at the ECMWF Copernicus Climate Data Store (CDS, [https://cds.climate.copernicus.eu/datasets/seasonal-original-single-levels?tab=overview](https:/cds.climate.copernicus.eu/datasets/seasonal-original-single-levels?tab=overview) ) and the NOAA NCEP CFS-model data archive ( [https://www.ncei.noaa.gov/data/climate-forecast-system/access/operational-9-month-forecast/](https:/www.ncei.noaa.gov/data/climate-forecast-system/access/operational-9-month-forecast) ), respectively. FuXi-ENS ensemble forecast data are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHudson, D., Alves, O., Hendon, H. H. \u0026amp; Marshall, A. G. Bridging the gap between weather and seasonal forecasting: intraseasonal forecasting for Australia. \u003cem\u003eQuart. J. Royal Meteoro Soc.\u003c/em\u003e \u003cb\u003e137\u003c/b\u003e, 673\u0026ndash;689 (2011).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRobertson, A. W., Kumar, A., Pe\u0026ntilde;a, M. \u0026amp; Vitart, F. Improving and Promoting Subseasonal to Seasonal Prediction. \u003cem\u003eBull. Am. Meteorol. Soc.\u003c/em\u003e \u003cb\u003e96\u003c/b\u003e, ES49\u0026ndash;ES53 (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVitart, F. et al. The Subseasonal to Seasonal (S2S) Prediction Project Database. \u003cem\u003eBull. Am. Meteorol. Soc.\u003c/em\u003e \u003cb\u003e98\u003c/b\u003e, 163\u0026ndash;173 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVitart, F. \u0026amp; Robertson, A. W. The sub-seasonal to seasonal prediction project (S2S) and the prediction of extreme events. \u003cem\u003enpj Clim. Atmos. Sci.\u003c/em\u003e \u003cb\u003e1\u003c/b\u003e, 3 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePortele, T. C. et al. Seasonal forecasts offer economic benefit for hydrological decision making in semi-arid regions. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 10581 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWhite, C. J. et al. Advances in the Application and Utility of Subseasonal-to-Seasonal Predictions. \u003cem\u003eBull. Am. Meteorol. Soc.\u003c/em\u003e \u003cb\u003e103\u003c/b\u003e, E1448\u0026ndash;E1472 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWhite, C. J. et al. Potential applications of subseasonal-to‐seasonal (S2S) predictions. \u003cem\u003eMeteorol. Appl.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e, 315\u0026ndash;325 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChevuturi, A. et al. Forecast skill of the Indian monsoon and its onset in the ECMWF seasonal forecasting system 5 (SEAS5). \u003cem\u003eClim. Dyn.\u003c/em\u003e \u003cb\u003e56\u003c/b\u003e, 2941\u0026ndash;2957 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCrespi, A., Petitta, M., Marson, P., Viel, C. \u0026amp; Grigis, L. Verification and Bias Adjustment of ECMWF SEAS5 Seasonal Forecasts over Europe for Climate Service Applications. \u003cem\u003eClimate\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 181 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDe Andrade, F. M. et al. Subseasonal Precipitation Prediction for Africa: Forecast Evaluation and Sources of Predictability. \u003cem\u003eWeather Forecast.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e, 265\u0026ndash;284 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeFlorio, M. J. et al. Experimental Subseasonal-to‐Seasonal (S2S) Forecasting of Atmospheric Rivers Over the Western United States. \u003cem\u003eJGR Atmos.\u003c/em\u003e \u003cb\u003e124\u003c/b\u003e, 11242\u0026ndash;11265 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEhsan, M. A. et al. Seasonal predictability of Ethiopian Kiremt rainfall and forecast skill of ECMWF\u0026rsquo;s SEAS5 model. \u003cem\u003eClim. Dyn.\u003c/em\u003e \u003cb\u003e57\u003c/b\u003e, 3075\u0026ndash;3091 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFerreira, G. W. S., Reboita, M. S. \u0026amp; Drumond, A. Evaluation of ECMWF-SEAS5 Seasonal Temperature and Precipitation Predictions over South America. \u003cem\u003eClimate\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 128 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSiegmund, J., Bliefernicht, J., Laux, P. \u0026amp; Kunstmann, H. Toward a seasonal precipitation prediction system for West Africa: Performance of CFSv2 and high-resolution dynamical downscaling. \u003cem\u003eJGR Atmos.\u003c/em\u003e \u003cb\u003e120\u003c/b\u003e, 7316\u0026ndash;7339 (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTian, D., Wood, E. F. \u0026amp; Yuan, X. CFSv2-based sub-seasonal precipitation and temperature forecast skill over the contiguous United States. \u003cem\u003eHydrol. Earth Syst. Sci.\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 1477\u0026ndash;1490 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, Q. J. et al. An evaluation of ECMWF SEAS5 seasonal climate forecasts for Australia using a new forecast calibration algorithm. \u003cem\u003eEnviron. Model. Softw.\u003c/em\u003e \u003cb\u003e122\u003c/b\u003e, 104550 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWeber, N. J. \u0026amp; Mass, C. F. Evaluating CFSv2 Subseasonal Forecast Skill with an Emphasis on Tropical Convection. \u003cem\u003eMon. Weather Rev.\u003c/em\u003e \u003cb\u003e145\u003c/b\u003e, 3795\u0026ndash;3815 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMandal, R. et al. Real time extended range prediction of heat waves over India. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 9008 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKueh, M. T. \u0026amp; Lin, C. Y. The 2018 summer heatwaves over northwestern Europe and its extended-range prediction. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 19283 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, O. \u0026amp; Lee, W. S. Enhancing seasonal predictability of the East Asian summer monsoon via optimized multi-model ensembles. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 24912 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHa, S., Wang, Q. J., Hur, J. \u0026amp; Im, E. S. Evaluating interannual variability in 1-month CFSv2 forecasts with different lead times at monthly and daily scales in South Korea. \u003cem\u003eEnviron. Res. Commun.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 041013 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIm, E. S. et al. An Evaluation of Temperature-Based Agricultural Indices Over Korea From the High-Resolution WRF Simulation. \u003cem\u003eFront. Earth Sci.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 656787 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee, Y., Pianosi, F., Pe\u0026ntilde;uela, A. \u0026amp; Rico-Ramirez, M. A. Skill of seasonal flow forecasts at catchment scale: an assessment across South Korea. \u003cem\u003eHydrol. Earth Syst. Sci.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e, 3261\u0026ndash;3279 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee, Y., Pe\u0026ntilde;uela, A., Pianosi, F. \u0026amp; Rico-Ramirez, M. A. Catchment‐scale skill assessment of seasonal precipitation forecasts across South Korea. \u003cem\u003eIntl J. Climatology\u003c/em\u003e. \u003cb\u003e43\u003c/b\u003e, 5092\u0026ndash;5111 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh, J., Oh, J. \u0026amp; Huh, M. A. Case Study of Bias Correction and the Dynamical Downscaling of CFSv2 S2S Forecasts Using a WRF Model: Heatwave in 2018 over South Korea. \u003cem\u003eAtmosphere\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 2086 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, L. et al. A machine learning model that outperforms conventional global subseasonal forecast models. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 6425 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, L. et al. FuXi: a cascade machine learning forecasting system for 15-day global weather forecast. \u003cem\u003enpj Clim. Atmos. Sci.\u003c/em\u003e \u003cb\u003e6\u003c/b\u003e, 190 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, D. H., Moon, I. J. \u0026amp; Joh, M. Recent Advances in AI-based Atmospheric General Circulation Models. \u003cem\u003eAtmosphere\u003c/em\u003e \u003cb\u003e35\u003c/b\u003e, 263\u0026ndash;276 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, Y., Lu, G., Wu, Z., He, H. \u0026amp; He, J. High-Resolution Dynamical Downscaling of Seasonal Precipitation Forecasts for the Hanjiang Basin in China Using the Weather Research and Forecasting Model. \u003cem\u003eJ. Appl. Meteorol. Climatology\u003c/em\u003e. \u003cb\u003e56\u003c/b\u003e, 1515\u0026ndash;1536 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhong, X. et al. FuXi-ENS: A machine learning model for medium-range ensemble weather forecasting. \u003cem\u003ePreprint at.\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2405.05925\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2405.05925\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHwang, J., Orenstein, P., Cohen, J., Pfeiffer, K. \u0026amp; Mackey, L. Improving Subseasonal Forecasting in the Western U.S. with Machine Learning. in \u003cem\u003eProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \u0026amp; Data Mining\u003c/em\u003e 2325\u0026ndash;2335 (ACM, Anchorage AK USA, (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/3292500.3330674\u003c/span\u003e\u003cspan address=\"10.1145/3292500.3330674\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKenne, A. D., Toure, M., Logamou Seknewna, L. \u0026amp; Ketsemen, H. L. Subseasonal Prediction of Summer Temperature in West Africa Using Artificial Intelligence: A Case Study of Senegal. \u003cem\u003eInt. J. Intell. Syst.\u003c/em\u003e \u003cb\u003e2024\u003c/b\u003e, 1\u0026ndash;19 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, C. et al. Improving the Accuracy of Subseasonal Forecasting of China Precipitation With a Machine Learning Approach. \u003cem\u003eFront. Earth Sci.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 659310 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, L. et al. Improving Subseasonal-to-Seasonal forecasts in predicting the occurrence of extreme precipitation events over the contiguous U.S. using machine learning models. \u003cem\u003eAtmos. Res.\u003c/em\u003e \u003cb\u003e281\u003c/b\u003e, 106502 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePinheiro, E. \u0026amp; Ouarda, T. B. M. J. Short-lead seasonal precipitation forecast in northeastern Brazil using an ensemble of artificial neural networks. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 20429 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, Y. I., Kim, D. \u0026amp; Lee, S. O. Prediction of Temperature and Heat Wave Occurrence for Summer Season Using Machine Learning. \u003cem\u003eJ. Korean Soc. Disaster Secur.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 27\u0026ndash;38 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh, S. G. et al. Deep learning model for heavy rainfall nowcasting in South Korea. \u003cem\u003eWeather Clim. Extremes\u003c/em\u003e. \u003cb\u003e44\u003c/b\u003e, 100652 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh, S. G. et al. Evaluation of Deep-Learning-Based Very Short-Term Rainfall Forecasts in South Korea. \u003cem\u003eAsia-Pac J. Atmos. Sci.\u003c/em\u003e \u003cb\u003e59\u003c/b\u003e, 239\u0026ndash;255 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCopernicus Climate Change Service. Seasonal forecast daily and subdaily data on single levels. ECMWF (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.24381/CDS.181D637E\u003c/span\u003e\u003cspan address=\"10.24381/CDS.181D637E\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSaha, S. et al. The NCEP Climate Forecast System Version 2. \u003cem\u003eJ. Clim.\u003c/em\u003e \u003cb\u003e27\u003c/b\u003e, 2185\u0026ndash;2208 (2014).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGubler, S. et al. Assessment of ECMWF SEAS5 Seasonal Forecast Performance over South America. \u003cem\u003eWeather Forecast.\u003c/em\u003e \u003cb\u003e35\u003c/b\u003e, 561\u0026ndash;584 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQiu, L., Im, E. S., Hur, J. \u0026amp; Shim, K. M. Added value of very high resolution climate simulations over South Korea using WRF modeling system. \u003cem\u003eClim. Dyn.\u003c/em\u003e \u003cb\u003e54\u003c/b\u003e, 173\u0026ndash;189 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHong, S. Y., Dudhia, J. \u0026amp; Chen, S. H. A Revised Approach to Ice Microphysical Processes for the Bulk Parameterization of Clouds and Precipitation. \u003cem\u003eMon Wea Rev.\u003c/em\u003e \u003cb\u003e132\u003c/b\u003e, 103\u0026ndash;120 (2004).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIacono, M. J. et al. Radiative forcing by long-lived greenhouse gases: Calculations with the AER radiative transfer models. \u003cem\u003eJ. Geophys. Res.\u003c/em\u003e \u003cb\u003e113\u003c/b\u003e, 2008JD009944 (2008).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJim\u0026eacute;nez, P. A. et al. A Revised Scheme for the WRF Surface Layer Formulation. \u003cem\u003eMon. Weather Rev.\u003c/em\u003e \u003cb\u003e140\u003c/b\u003e, 898\u0026ndash;918 (2012).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, F. \u0026amp; Dudhia, J. Coupling an Advanced Land Surface\u0026ndash;Hydrology Model with the Penn State\u0026ndash;NCAR MM5 Modeling System. Part I: Model Implementation and Sensitivity. \u003cem\u003eMon Wea Rev.\u003c/em\u003e \u003cb\u003e129\u003c/b\u003e, 569\u0026ndash;585 (2001).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHong, S. Y., Noh, Y. \u0026amp; Dudhia, J. A. New Vertical Diffusion Package with an Explicit Treatment of Entrainment Processes. \u003cem\u003eMon. Weather Rev.\u003c/em\u003e \u003cb\u003e134\u003c/b\u003e, 2318\u0026ndash;2341 (2006).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKain, J. S. The Kain\u0026ndash;Fritsch Convective Parameterization: An Update. \u003cem\u003eJ. Appl. Meteor.\u003c/em\u003e \u003cb\u003e43\u003c/b\u003e, 170\u0026ndash;181 (2004).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMa, L. M. \u0026amp; Tan, Z. M. Improving the behavior of the cumulus parameterization for tropical cyclone prediction: Convection trigger. \u003cem\u003eAtmos. Res.\u003c/em\u003e \u003cb\u003e92\u003c/b\u003e, 190\u0026ndash;211 (2009).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7639295/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7639295/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDespite the growing availability of global ensemble forecasts, their application at regional scales remains challenging due to the coarse spatial resolution. Meanwhile, machine learning (ML)-based forecasts are emerging as promising alternatives to traditional physics-based models, yet they are mostly deterministic, and their potential has not been fully explored beyond the medium-range timeframe. This study investigates the potential of integrating global ML-based ensemble forecasts (FuXi-ENS) with dynamical downscaling to improve one-month temperature prediction over South Korea. The forecasting performance of FuXi-ENS, in terms of both temporal and spatial patterns, is compared against state-of-the-art physics-based model forecasting data from NOAA and ECMWF, which serve as benchmarks. The superiority of FuXi-ENS becomes pronounced after dynamical downscaling, highlighting the added value of ML-based forecasts when combined with high-resolution physical modeling. Overall, this study offers a comprehensive assessment of extended-range temperature prediction over South Korea, illustrating the operational potential of hybrid approaches that combine global ML models with regional dynamical downscaling and providing insights for the future development of hybrid forecasting systems.\u003c/p\u003e","manuscriptTitle":"Enhancing 1-month temperature predictions in South Korea through dynamical downscaling of machine learning global ensemble forecasts","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-07 09:05:44","doi":"10.21203/rs.3.rs-7639295/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-30T02:32:20+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-13T02:16:29+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-09T06:50:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"26926177227105619316578071104303820122","date":"2025-11-11T14:06:32+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"58272669225643157104391047918940895947","date":"2025-11-10T00:05:03+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"78586086502132788096168016666030256052","date":"2025-10-16T12:49:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"78586086502132788096168016666030256052","date":"2025-09-25T11:08:54+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-24T19:11:28+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-09-24T05:38:53+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-23T05:20:31+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-20T01:27:33+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-09-17T10:49:54+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"6335d429-ed92-4d24-9831-4dbdaed8a792","owner":[],"postedDate":"October 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":55774830,"name":"Earth and environmental sciences/Climate sciences"},{"id":55774831,"name":"Earth and environmental sciences/Environmental sciences"}],"tags":[],"updatedAt":"2026-03-01T05:23:41+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-07 09:05:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7639295","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7639295","identity":"rs-7639295","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}