Subseasonal Representation and Predictability of North American Weather Regimes Using Cluster Analysis

Maria J. Molina aUniversity of Maryland, College Park, College Park, Maryland
bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Maria J. Molina in
Current site
Google Scholar
PubMed
Close
https://orcid.org/0000-0001-8539-8916
,
Jadwiga H. Richter bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Jadwiga H. Richter in
Current site
Google Scholar
PubMed
Close
,
Anne A. Glanville bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Anne A. Glanville in
Current site
Google Scholar
PubMed
Close
,
Katherine Dagon bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Katherine Dagon in
Current site
Google Scholar
PubMed
Close
,
Judith Berner bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Judith Berner in
Current site
Google Scholar
PubMed
Close
,
Aixue Hu bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Aixue Hu in
Current site
Google Scholar
PubMed
Close
, and
Gerald A. Meehl bNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Gerald A. Meehl in
Current site
Google Scholar
PubMed
Close
Open access

Abstract

This study focuses on assessing the representation and predictability of North American weather regimes, which are persistent large-scale atmospheric patterns, in a set of initialized subseasonal reforecasts created using the Community Earth System Model, version 2 (CESM2). The k-means clustering was used to extract four key North American (10°–70°N, 150°–40°W) weather regimes within ERA5 reanalysis, which were used to interpret CESM2 subseasonal forecast performance. Results show that CESM2 can recreate the climatology of the four main North American weather regimes with skill but exhibits biases during later lead times with overoccurrence of the West Coast high regime and underoccurrence of the Greenland high and Alaskan ridge regimes. Overall, the West Coast high and Pacific trough regimes exhibited higher predictability within CESM2, partly related to El Niño. Despite biases, several reforecasts were skillful and exhibited high predictability during later lead times, which could be partly attributed to skillful representation of the atmosphere from the tropics to extratropics upstream of North America. The high predictability at the subseasonal time scale of these case-study examples was manifested as an “ensemble realignment,” in which most ensemble members agreed on a prediction despite ensemble trajectory dispersion during earlier lead times. Weather regimes were also shown to project distinct temperature and precipitation anomalies across North America that largely agree with observational products. This study further demonstrates that unsupervised learning methods can be used to uncover sources and limits of subseasonal predictability, along with systematic biases present in numerical prediction systems.

Significance Statement

North American weather regimes are large-scale atmospheric patterns that can persist for several days. Their skillful subseasonal (2 weeks or greater) prediction can provide valuable lead time to prepare for temperature and precipitation anomalies that can stress energy and water resources. The purpose of this study was to assess the climatological representation and subseasonal predictability of four key North American weather regimes using a research subseasonal prediction system and clustering analysis. We found that the Pacific trough and West Coast high regimes exhibited higher predictability than other regimes and that skillful representation of conditions across the tropics and extratropics can increase predictability during later lead times. Future work will quantify causal pathways associated with high predictability.

© 2023 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Maria J. Molina, mjmolina@umd.edu

Abstract

This study focuses on assessing the representation and predictability of North American weather regimes, which are persistent large-scale atmospheric patterns, in a set of initialized subseasonal reforecasts created using the Community Earth System Model, version 2 (CESM2). The k-means clustering was used to extract four key North American (10°–70°N, 150°–40°W) weather regimes within ERA5 reanalysis, which were used to interpret CESM2 subseasonal forecast performance. Results show that CESM2 can recreate the climatology of the four main North American weather regimes with skill but exhibits biases during later lead times with overoccurrence of the West Coast high regime and underoccurrence of the Greenland high and Alaskan ridge regimes. Overall, the West Coast high and Pacific trough regimes exhibited higher predictability within CESM2, partly related to El Niño. Despite biases, several reforecasts were skillful and exhibited high predictability during later lead times, which could be partly attributed to skillful representation of the atmosphere from the tropics to extratropics upstream of North America. The high predictability at the subseasonal time scale of these case-study examples was manifested as an “ensemble realignment,” in which most ensemble members agreed on a prediction despite ensemble trajectory dispersion during earlier lead times. Weather regimes were also shown to project distinct temperature and precipitation anomalies across North America that largely agree with observational products. This study further demonstrates that unsupervised learning methods can be used to uncover sources and limits of subseasonal predictability, along with systematic biases present in numerical prediction systems.

Significance Statement

North American weather regimes are large-scale atmospheric patterns that can persist for several days. Their skillful subseasonal (2 weeks or greater) prediction can provide valuable lead time to prepare for temperature and precipitation anomalies that can stress energy and water resources. The purpose of this study was to assess the climatological representation and subseasonal predictability of four key North American weather regimes using a research subseasonal prediction system and clustering analysis. We found that the Pacific trough and West Coast high regimes exhibited higher predictability than other regimes and that skillful representation of conditions across the tropics and extratropics can increase predictability during later lead times. Future work will quantify causal pathways associated with high predictability.

© 2023 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Maria J. Molina, mjmolina@umd.edu

1. Introduction

During the 1960s, E. N. Lorenz introduced the concept of sensitivity to initial state for numerical weather prediction (Lorenz 1963). The idea became popularly known as chaos theory or the “butterfly effect,” in which small perturbations to the initial state can yield substantial trajectory (i.e., state estimate) deviations over time. Chaos theory helps to explain why the skillful deterministic prediction of weather at the subseasonal time scale (from 2 weeks to 2 months) using state-of-the-art forecasting models remains extremely challenging (Barnes et al. 2020).

While the inherent limit to predictability of the atmosphere cannot be changed, running an ensemble of simulations allows representation of such forecast uncertainty in weather and climate forecasts. Small random perturbations to the initial state can be used to create ensembles for weather forecasting (Murphy 1988; Buizza et al. 2005; Bauer et al. 2015; Zhou et al. 2017), subseasonal-to-decadal prediction (Meehl et al. 2016, 2021, 2022; Vitart et al. 2017; Yeager et al. 2018, 2022; Pegion et al. 2019; Richter et al. 2020, 2022), and climate projections (Deser et al. 2020; Maher et al. 2021). The variance across the ensemble members (also called spread) can provide an indication of forecast uncertainty and predictability (Albers and Newman 2019). Reduced ensemble spread has been attributed to initial or boundary conditions that offer added predictability in the Earth system. Such forecasts are often referred to as “forecasts of opportunity” if skillful (Mariotti et al. 2020). Thus, sensitivity to initial state also means that skillful prediction is easier when starting from certain atmospheric flow configurations than from others (Ferranti et al. 2015).

Large-scale atmospheric flow configurations that are quasi-stationary, persistent, and/or recurrent are often referred to as “weather regimes” (WRs) and are based on the idea that planetary-scale patterns or states can influence the behavior of transient synoptic-scale disturbances (Reinhold and Pierrehumbert 1982; Michelangeli et al. 1995). Weather regimes help describe large-amplitude flow patterns within lower temporal frequencies, which extend beyond the lifetime of individual weather disturbances and appear repeatedly across certain geographic locations in the midlatitudes (Muñoz et al. 2017; Lee et al. 2019; Robertson et al. 2020). While regimes typically manifest as persistent and recurrent patterns, studies rooted in dynamical systems theory view regimes as special states, namely ghosts of attractor basins (Kimoto and Ghil 1993a,b). Berner and Branstator (2007) demonstrate how overlapping Gaussian regime density distributions (i.e., preferred “states”) can result in preferred weather patterns. Weather regimes also have an imprint on weather at the surface, with certain regimes providing favorable large-scale environmental conditions for notable temperature and precipitation anomalies, along with impacts to storm tracks, moisture flux, and atmospheric rivers (Robertson and Ghil 1999; Amini and Straus 2019).

Since weather regimes typically persist from several days to 2 weeks, they highlight a physically meaningful subseasonal component of variability (Hochman et al. 2021) and have been used by numerous studies to assess subseasonal predictability (e.g., Cassou 2008; Ferranti et al. 2015; Muñoz et al. 2017; Ferranti et al. 2018; Huang et al. 2020; Robertson et al. 2020; Büeler et al. 2021; Cortesi et al. 2021). In this study, we purposefully distinguish weather regimes from “weather typing” or “weather types,” which generally refer to the classification of synoptic-scale weather patterns on time scales shorter than the subseasonal and may be embedded within weather regimes (Lamb 1972; Yarnal 1993; Conway and Jones 1998; Prein et al. 2019). Weather regimes have also been referred to as “circulation regimes” (e.g., Straus et al. 2007), but we refer to them as weather regimes throughout this article for consistency.

The inception of the concept of weather regimes saw the rise of the use of data compositing and spectral methods for their characterization (e.g., Rex 1951; Reinhold and Pierrehumbert 1982). More recent studies have focused on the use of unsupervised learning, such as k-means clustering or Gaussian mixture models, for extracting weather regimes from large datasets (e.g., Kimoto and Ghil 1993a,b; Berner and Branstator 2007; Smyth et al. 1999; Vitart 2014). Unsupervised learning, which refers to characterization of unlabeled data for pattern discovery (Celebi and Aydin 2016), is in itself not a new field of study nor is k-means clustering a new tool (MacQueen 1967). However, large and high-quality datasets focusing on subseasonal prediction have grown substantially in number over the last decade (Vitart 2017; Pegion et al. 2019; Richter et al. 2020, 2022). Assessing the climatological representation of weather regimes within these new initialized prediction datasets, along with their imprint to surface weather anomalies, remains of high importance for model validation.

The continued growing interest in subseasonal prediction across scientific sectors (e.g., government operations, research, and private industry) is a testament to the societal need for a deeper understanding of predictive capabilities at such time scales (Merryfield et al. 2020; Meehl et al. 2021). The predictability of North American weather regimes at subseasonal time scales within certain initialized prediction systems [e.g., European Centre for Medium-Range Weather Forecasts (ECMWF) and National Centers for Environmental Prediction (NCEP) Climate Forecast System, version 2 (CFSv2)] has been recently assessed using k-means clustering (Vigaud et al. 2018; Robertson et al. 2020). Here we apply this framework to subseasonal forecasts created using the Community Earth System Model, version 2 (CESM2; Danabasoglu et al. 2020; Richter et al. 2022), which is a new research prediction system contributing to the Subseasonal Experiment (SubX; Pegion et al. 2019).

Subseasonal prediction remains particularly challenging because the sources of predictability at such time scales are limited. Predictability stemming from atmospheric initial conditions is substantially reduced beyond approximately 2 weeks, and the ocean generally does not offer added predictability until a trajectory reaches the seasonal time scale (Meehl et al. 2021). Further motivating our study is that weekly real-time CESM2 forecasts (Richter et al. 2022) are being contributed to a multimodel mean ensemble used to issue experimental NOAA subseasonal outlooks as part of the SubX experiment (Pegion et al. 2019) making this an opportune time to assess the capabilities of the prediction system and its sources and limits of predictability. By focusing on predictability of weather regimes within initialized coupled forecasts consisting of 11 ensemble members, we will also explore ensemble spread as a function of lead time.

2. Methods

a. Numerical model configuration

The 11-member ensemble of subseasonal initialized reforecasts used in this study was created using CESM2, which is an Earth system model developed at the National Center for Atmospheric Research in collaboration with the broader scientific and academic community (Danabasoglu et al. 2020). CESM2 is composed of numerous coupled Earth system model components, which include the Community Atmosphere Model, version 6.0 (CAM6); Parallel Ocean Program, version 2 (POP2; Smith et al. 2010; Danabasoglu et al. 2012, 2020); Marine Biogeochemistry Library (MARBL; Moore et al. 2002, 2004, 2013); NOAA WaveWatch-III ocean surface wave prediction model, version 3.14 (Tolman 2009); Community Ice Code, version 5.1.2 (CICE5; Hunke et al. 2015); and Community Land Model, version 5 (CLM5; Lawrence et al. 2019). CAM6 was run with nominal 1° horizontal resolution and 32 vertical levels up to a height of approximately 2 hPa. POP2 was run with 60 vertical levels within the upper 160 m of the ocean, along with a nominal 1° horizontal resolution, which is uniform in the zonal direction at 1.125° but varies across the meridional direction from 0.27° at the equator to 0.64° at the midlatitudes.

CESM2 subseasonal reforecasts were initialized weekly every Monday (1999–2019) and were carried out following SubX protocol (Pegion et al. 2019). The atmosphere was initialized using the NCEP CFSv2 operational model (Saha et al. 2006). Ocean and sea ice initial conditions originate from a reforecast ocean–sea ice coupled configuration of CESM2 (CAM6) forced with a surface–atmospheric dataset based on the Japanese 55-year Reanalysis driving ocean product (JRA55-do forcing; Tsujino et al. 2018; Yeager et al. 2018). Land initial conditions were created using the stand-alone CLM5 with a 700-yr spinup that includes biogeochemistry-driven crops and glacial observations. The random field perturbation method (Magnusson et al. 2009; Richter et al. 2020) was applied at the initial time to the atmospheric initial conditions to generate the 11-member ensemble. A complete description of the CESM2 subseasonal prediction configuration is available in Richter et al. (2022).

b. Data

Several observation and reanalysis products were used for comparison with the CESM2 subseasonal reforecasts and include the fifth major global reanalysis produced by ECMWF (ERA5; Hersbach et al. 2020), the NOAA Climate Prediction Center (CPC) global daily gridded surface air temperature, CPC global unified gauge-based analysis of daily precipitation (Xie et al. 2007), and the NOAA Global Precipitation Climatology Project (GPCP; Adler et al. 2003). Following SubX protocol (Pegion et al. 2019), all observation, reanalysis, and CESM2 data were interpolated onto a 1° global grid. Daily averages were computed for 500-hPa geopotential height (m) and 2-m temperature (°C). The daily sum was computed for precipitation (mm day−1). Lead-time bias-corrected anomalies were used for all variables to remain consistent with SubX protocol (Pegion et al. 2019) and were calculated from a daily climatology that spans 1999–2019 (see appendix A for more details).

We note that using the full climate period in anomaly calculations can introduce artificial skill for assessing model performance, given that future years are not available in real time (Risbey et al. 2021; Meehl et al. 2022). However, generating reforecasts with CESM2 is a computationally intensive task and we are therefore limited to a 20-yr climatological period. For consistency, the same climatology period (1999–2019) was used for ERA5, NOAA CPC, and NOAA GPCP. A 5-day running average was also applied along the lead-time dimension of 500-hPa geopotential height anomalies to filter out variability on the weather time scale, an approach commonly used to emphasize the low-frequency component of variability (Straus et al. 2007).

c. Weather regimes

Numerous studies have devised variations to the earlier introduced methods (Rex 1951; Reinhold and Pierrehumbert 1982) for extracting weather regimes from data (e.g., Vitart 2014). However, given that the objective of this study is to assess both the representation of weather regimes and model performance of CESM2, we follow the methodology of Vigaud et al. (2018) and Robertson et al. (2020) closely. The steps for computation of weather regimes include (i) extract the 12 leading principal components (PCs) from ERA5 500-hPa geopotential height low-pass-filtered daily anomalies (1999–2019) for dimensionality reduction (which explain 92.8% of the variance), (ii) identify weather regimes within the 12 leading PCs from ERA5 using k-means clustering (Michelangeli et al. 1995), and (iii) project CESM2 500-hPa geopotential height daily anomalies that were lead-time bias corrected and low-pass-filtered (1999–2019) onto the ERA5 clusters. The 12 leading PCs were extracted from low-pass-filtered data in order to further emphasize the large-scale subseasonal component of variability and filter out faster synoptic scales (Robertson et al. 2020). Extraction of the 12 leading PCs also reduces the dimensionality of the clustering problem (from over 6500 input data points to just 12) and limits linear dependence among the training data.

The k-means clustering algorithm was trained using four clusters (i.e., weather regimes) from the 12 leading PCs extracted from ERA5. Given that model validation is an objective of this study, and weather regimes exhibit nonstationarity in a warming climate (Kageyama et al. 1999; Li et al. 2012; Bruyère et al. 2017; Steinschneider et al. 2019), we train k-means using ERA5 from the same time period as CESM2 (1999–2019). We further note that training a machine learning model using different climate periods (without overlapping years between the training and testing sets) would be important in a supervised learning setting, where a machine learning model would be tasked with predicting weather regimes. The focus here is instead on using unsupervised learning for signal extraction (weather regimes) in order to validate CESM2 during earlier lead times and then assess its performance during later lead times at the subseasonal time scale. We found that ERA5 weather regimes were consistent with those identified by Vigaud et al. (2018), Lee et al. (2019), and Robertson et al. (2020) over a similar time period. Performing k-means clustering using k-fold cross validation (k = 5) or a larger number of clusters (e.g., 5 and 6) did not change the predominant patterns of the four identified weather regimes, providing further evidence that the four clusters represent a meaningful component of subseasonal variability and are the most commonly occurring patterns. The training hyperparameters for the k-means clustering algorithm are available in Table 1 and were chosen following a hyperparameter grid search that involved varying the number of clusters, number of leading PCs, centroid initialization method, input months, North American domain, and training iterations. Additionally, an inertia metric was used to quantitatively assess the number of clusters to use for our study. The inertia metric is defined as the sum of squared distances between the samples in a cluster and the respective cluster centroid,
i=1nminμjC(xiμj2),
where n is the number of samples x, xi are the samples in the cluster, μj is the mean of the samples in the cluster (i.e., centroid), and C are the clusters. A smaller inertia metric is preferred, which would suggest that samples assigned to a cluster are more similar (within their respective cluster). Note that the use of other metrics or research goals may result in a different number of chosen regimes (e.g., Christiansen 2007; Riddle et al. 2013). Data used for training k-means was constrained to extended cool season months (October–March), when teleconnections from the tropics to extratropics tend to be stronger (Madden 1986; Mayer and Barnes 2021).
Table 1

Hyperparameters and assessment of the k-means clustering algorithm.

Table 1

d. Software

The Python language software for k-means was sourced from sklearn (Pedregosa et al. 2011) and the following libraries were used for data analysis and visualization: numpy (Harris et al. 2020), pandas (McKinney et al. 2011), xarray (Hoyer and Hamman 2017), matplotlib (Hunter 2007), cartopy (Met Office 2010), and scipy (Virtanen et al. 2020).

3. Results

a. Weather regime patterns and representation in CESM2

The four dominant weather regimes derived from ERA5 are shown in Fig. 1 as composites of daily 500-hPa geopotential height anomalies assigned to the nearest k-means cluster. While k-means was trained using data from October to March and across 10°–70°N and 150°–40°W, Fig. 1 shows data hemispherically in order to gain a broader perspective of geopotential height fields both upstream and downstream of North America. Some large-magnitude anomalies are located outside the training domain (e.g., Fig. 1c), but we emphasize that these areas were not used for training k-means given our focus on North America. ERA5 data contained in Fig. 1 is arranged as the CESM2 reforecasts, with weekly Monday starts that extend 6 weeks into the future for a more straightforward comparison with CESM2 weather regime patterns and frequencies. Since September initializations of CESM2 extend into October (e.g., weeks 5–6), ERA5 data from the month of September was also included in Fig. 1. For consistency with Lee et al. (2019) and Robertson et al. (2020), we refer to the four weather regimes using similar nomenclature: West Coast high (Fig. 1a), Pacific trough (Fig. 1b), Greenland high (Fig. 1c), and Alaskan ridge (Fig. 1d).

Fig. 1.
Fig. 1.

WR composites of daily 500-hPa geopotential height anomalies (m) computed from ERA5 (September–March 1999–2019). WRs are illustrated hemispherically, but 10°–70°N, 150°–40°W was used for their computation, as indicated with the black-outlined polygon. Hatching indicates statistical significance at the 99% confidence level using a two-tailed 10 000-member bootstrap resampling test (Chernick 2011).

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Over North America, the West Coast high, Pacific trough, and Greenland high regimes show meridionally oriented 500-hPa geopotential height anomalies (Figs. 1a–c), which resemble Rossby wave trains extending from the Pacific Ocean to the Atlantic Ocean, a feature that was also noted by Lee et al. (2019) and Robertson et al. (2020). The West Coast high regime consists of positive geopotential height anomalies across the west and east coasts of North America, whereas the Greenland high regime consists of negative 500-hPa geopotential height anomalies over both coasts (Figs. 1a,c). While not as robust as the pattern noted by Robertson et al. (2020), the Greenland high regime also contains anomalies that extend zonally into the Atlantic Ocean, resembling the negative phase of the North Atlantic Oscillation (NAO; Wallace and Gutzler 1981). A defining characteristic of the Pacific trough regime is a strong negative geopotential height anomaly over the North Pacific and Alaska (Fig. 1b). The Alaskan ridge regime can be identified by its comparatively more zonally oriented geopotential height anomalies over the contiguous United States (CONUS), but a Rossby wave train pattern is still evident extending from the North Pacific into the southern CONUS (Fig. 1d). The Greenland high and West Coast high regimes also resemble the Pacific–North American teleconnection pattern (positive and negative, respectively).

The West Coast high regime occurs with greater frequency1 (30.4%) than the Pacific trough (27.4%), Greenland high (22.8%), and Alaskan ridge (19.3%) regimes, similar to frequencies identified by Robertson et al. (2020) using MERRA (Rienecker et al. 2011). All weather regimes are highly persistent with little daily transition from one regime to another, with probabilities of 84%–87% that they will persist (Table 2; statistically significant at the 0.05% level using the chi-squared test). When considering initial weather regimes (the regime ongoing at the start of a reforecast period), regime duration is approximately 1 week, but with the caveat that duration was inherently truncated due to reforecast initialization. A key takeaway is that a transition to a different weather regime has likely occurred by weeks 3–4 (Fig. 2). However, we also note that there is substantial variance in regime duration; a few Pacific trough and West Coast high regime events persist for more than 30 days (Fig. 2; black bars). The probability for an initial weather regime to transition to another weather regime are also indicated in Table S1 of the online supplemental material, which shows that the Alaskan ridge regime has a 43% probability of transitioning to the West Coast high regime, but there is a lack of statistical significance that is likely partly due to the limited sample size in initial weather regime event transitions (543 unique reforecast start dates during the extended cool season).

Table 2

Contingency table of daily ERA5 weather regime transitions (October–March 1999–2019;), which show frequencies from one regime (i.e., chosen from one of the column labels) to another regime (i.e., chosen from one of the row labels). Daily frequency of persistence is also shown (e.g., from WR1 to WR1). A chi-squared statistical significance test rejects the null hypothesis that regimes are independent (0.05% level of a χ2 test). Probabilities are indicated within parentheses, computed using the column frequencies.

Table 2
Fig. 2.
Fig. 2.

Duration (days) of each initial (i.e., first) WR for all CESM2 initialized reforecasts (October–March 1999–2019). The blue bars show the CESM2 11-member ensemble mean, and the error bars indicate 11-member ensemble spread. ERA5 WR duration is also indicated for reference (black bars). Weeks 3–4 are indicated with gray shading.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

The individual CESM2 ensemble members (not the ensemble mean) and all lead times are then projected on the ERA5 EOFs, generating pseudoPCs, putting the model states in the same PC space as the ERA5 states. Then clusters are assigned to the model states based on the nearest ERA5 centroid to assess the occurrence and evolution of weather regimes within the subseasonal initialized prediction system (i.e., model validation). Similar patterns emerge for composites of 500-hPa geopotential height anomalies averaged over each weather regime (Fig. 3), as further evidenced with strong Pearson correlations (≥0.9) between ERA5 and CESM2 weather regimes. Rossby wave train patterns are evident for all weather regimes, with the Greenland high regime also displaying an NAO-like pattern across the Atlantic Ocean (Fig. 3). The weather regime frequencies and anomaly amplitudes within CESM2 are comparable to ERA5 (Fig. 3). All CESM2 weather regimes are also highly persistent with little daily transition from one regime to another, with probabilities of 84%–88% that they will persist. Initial weather regime duration is consistent with ERA5 (approximately 1 week; Fig. 2), with a few CESM2 weather regime events persisting longer than ERA5 events (Fig. 2). The probability for an initial weather regime to transition to another regime is also consistent with ERA5 (Table S1 in the online supplemental material).

Fig. 3.
Fig. 3.

As in Fig. 1, but for CESM2 daily 500-hPa geopotential height anomalies (m) (September–March 1999–2019). Individual ensemble members (11 total) and lead days 0–42 are considered.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Frequencies of each weather regime as a function of lead time are shown in Fig. 4. For easier comparison of weather regimes across data products, ERA5 is arranged as the CESM2 reforecasts with weekly Monday starts, which was also done for Fig. 1. At lead time zero (i.e., analysis), weather regime frequencies within CESM2 are similar to those in ERA5 for all four weather regimes (Fig. 4). By weeks 5–6, the frequency of the West Coast high regime increases by approximately 5%–10% as compared with its frequency during the analysis period (day 0). In contrast, the frequencies of the Alaskan ridge and Greenland high regimes decrease by approximately 5% by weeks 5–6 as compared with their frequencies during day 0 (Fig. 4). While the further development and persistence of the West Coast high regime appears to be a bias of subseasonal forecasts initialized with CESM2, overall, these results show that weather regimes are represented skillfully (Fig. 4), consistent with Simpson et al. (2020), which found that the representation of the Pacific jet stream, large-scale atmospheric circulation, and storm tracks during winter months are much improved in CESM2 as compared with CESM1 and other models. We propose that analysis of weather regime frequencies could be of use to model developers, as they offer insight into persistent patterns that can influence surface conditions. For example, Danabasoglu et al. (2020) found that there are warm biases during winter months in northern high latitudes.

Fig. 4.
Fig. 4.

Percentage of days of WR occurrence for CESM2 (11-member ensemble mean) and corresponding ERA5 data arranged as the CESM2 reforecasts (e.g., weeks 1–2). ERA5 and CESM2 analysis (lead day 0) are indicated in the far-left columns, and CESM2 weeks 1–2, 3–4, and 5–6 reforecasts are indicated in the following columns with corresponding ERA5. WRs are indicated in the legend. Error bars represent spread among individual CESM2 ensemble members.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

b. Climatological representation of weather regime imprint on surface weather

To assess the imprint of weather regimes on surface weather, temperature and precipitation anomalies associated with the four weather regimes are shown across North America (Figs. 5 and 6). Given that ERA5 is a model derived product, we also show temperature and precipitation anomalies computed from observations often used for subseasonal forecast verification by NOAA CPC. Here we show weeks 3–4 anomalies for CESM2 to focus on representation skill at the subseasonal time scale (Figs. 5 and 6), but we also consider weeks 1–2 to help control for systematic biases that may emerge within CESM2, and weeks 5–6 to explore the upper limits of the prediction system (Figs. S1–S4 in the online supplemental material). As before, data from observations and ERA5 are arranged in a similar format as CESM2 (e.g., weekly Monday starts during 1999–2019).

Fig. 5.
Fig. 5.

Mean 2-m temperature anomalies (°C) for each WR during September–March 1999–2019 for (left) NOAA CPC data, (center) ERA5, and (right) weeks 3–4 CESM2 reforecasts. Hatching indicates (two tailed) statistical significance at the 99% confidence level using a 10 000-member bootstrap. ACC between NOAA CPC and ERA5 (land only) is shown in (b), (e), (h), and (k), and ACC between ERA5 and CESM2 (land only) is shown in (c), (f), (i), and (l). The insets show Hawaii.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Fig. 6.
Fig. 6.

As in Fig. 5, but for mean daily precipitation anomalies (mm day−1).

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Weeks 3–4 temperature and precipitation anomalies associated with specific weather regimes are relatively consistent across NOAA CPC and ERA5 (Figs. 5 and 6), with the exception of a few very localized anomalies. These exceptions include comparably warmer anomalies across the Northwest region of North America in ERA5 for the West Coast high regime (Fig. 5b) and comparably drier anomalies across Florida in ERA5 for the Alaskan ridge regime (Fig. 6k), with both examples being statistically significant at the 99% confidence level.

Large-scale spatial anomaly patterns are relatively similar between CESM2 weeks 1 and 2 and observations across all weather regimes, but CESM2 anomaly patterns are also smoother, with fewer sharp contrasts between anomaly magnitudes across neighboring grid cells (Figs. S1 and S2 in the online supplemental material). This anomaly artifact is likely partly attributable to topographical features (e.g., mountain ranges) being smoothed out within the model fields or localized extremes potentially associated with convective events and mesoscale features, suggesting that CESM2 reforecast anomalies associated with specific weather regimes will be less useful for highly localized extreme events (at least at this 1° grid spacing). We also note that temperature and precipitation anomalies from CESM2 stratified by weather regimes during weeks 5 and 6 are less representative of observations due to smaller anomaly magnitudes (Figs. S3 and S4 in the online supplemental material), likely due in part to ensemble spread. CESM2 weeks 3 and 4 generally show similar spatial patterns of temperature and precipitation anomalies for all weather regimes as compared with NOAA CPC and ERA5, albeit with comparatively weaker anomaly magnitudes contained within CESM2 (Figs. 5 and 6). CESM2 also exhibits skillful representation of precipitation anomalies along coastal areas, particularly the West Coast of the United States, but anomalies remain comparatively muted across the continental interior (Figs. 5 and 6).

Figures 5 and 6 show that weather regimes are associated with distinct temperature and precipitation anomalies (corroborating past work, e.g., Rojas et al. 2013; Zampieri et al. 2017; Barthlott and Hoose 2018). The West Coast high regime shows negative temperature anomalies across the central United States and Intermountain West, associated with persistent troughing and cold air advection (Figs. 5a–c and 6a–c; statistically significant). Positive temperature anomalies are also evident across the East Coast and western Canada, and negative precipitation anomalies over the West Coast, related to persistent high pressure limiting precipitation. Moisture is also transported northward on the western edge of persistent high pressure, contributing to positive precipitation anomalies over the eastern United States, Alaska, and western Canada (Figs. 5a–c and 6a–c; statistically significant). The Pacific trough regime is characterized by positive temperature anomalies over the northern United States and central Canada, which is related to the ridge east of the trough (Figs. 5d–f and 6d–f; statistically significant). Positive precipitation anomalies are also evident over the western United States, largely driven by onshore flow on the south side of a persistent area of low pressure over the North Pacific, consistent with results shown by Vigaud et al. (2018). Persistent onshore flow on the north side of an area of low pressure over the southeastern United States also contributes to positive precipitation anomalies over the East Coast (Figs. 5d–f and 6d–f; statistically significant). In contrast, negative precipitation anomalies over western Alaska are driven by continental air being advected westward around a persistent area of low pressure (Figs. 5d–f and 6d–f; statistically significant).

The Greenland high regime exhibits negative temperature and precipitation anomalies across the eastern United States, related to southward advection of cold and dry air on the backside of intensified low pressure (Figs. 5g–i and 6g–i; statistically significant). This anomaly pattern is similar to the negative phase of the NAO (Hurrell et al. 2001). The Greenland high regime also has positive temperature anomalies across the high plains and Intermountain West related to a persistent ridge (Figs. 5g–i and 6g–i; statistically significant). The Alaskan ridge regime is characterized by a persistent ridge over Alaska and the clockwise flow drives positive temperature anomalies over western Alaska and negative temperature anomalies over the upper Midwest of the United States and central and western Canada, consistent with Vigaud et al. (2018). This persistent ridge also contributes to offshore flow and negative precipitation anomalies over British Columbia and positive precipitation anomalies over western Alaska (Figs. 5j–l and 6j–l; statistically significant). The Alaskan ridge regime also contains positive temperature anomalies and negative precipitation anomalies over the southeastern United States related to anomalous high pressure (Figs. 5j–l and 6j–l; statistically significant). Positive precipitation anomalies over the southwestern United States are related to persistent onshore flow of moisture on the western edge of anomalous high pressure over the southern United States.

c. Prediction skill of weather regimes within CESM2

Here we shift focus from model validation to forecast verification and assess CESM2 prediction skill for weather regimes. We first compute anomaly correlation coefficient (ACC) for each lead day using
ACC(d)=i=1n(d)(oio¯)(fif¯)i=1n(d)(oio¯)2i=1n(d)(fif¯)2,
where n(d) is the number of samples for a lead day d, oi are observations, and fi are ensemble mean forecasts for the respective lead day; o¯ and f¯ indicate the mean for the respective lead day. Each observation or forecast is one of the four weather regime classes (class label of 1, 2, 3, or 4) computed from the preprocessed 500-hPa geopotential height anomalies. Mean square error (MSE) for each lead day is
MSE(d)=1n(d)i=1n(d)(oifi)2,
where n(d), oi, and fi are the number of samples, observations, and ensemble mean forecasts.

ACC(d) between CESM2 and ERA5 shows strong agreement among weather regime occurrence within the first 10 lead days, generally exceeding 0.4 (Fig. 7a). At longer lead times, ACC(d) decreases from approximately 0.2 by week 2, to approximately 0.1 by week 3, and 0.0 by week 5 (Fig. 7a). MSE(d) between CESM2 and ERA5 weather regime occurrence shows a similar trend; MSE increases with increasing lead time (Fig. 7b). The largest increase in MSE occurs within the first 10 days and it continues to increase during later lead times, but at a slower rate (MSE is approximately 2 at 2 weeks, which increases to 2.5 at 5 weeks and beyond; Fig. 7b).

Fig. 7.
Fig. 7.

(a) ACC(d) and (b) MSE(d) for CESM2 (11-member ensemble mean; 1999–2019) WRs as a function of lead (days) during September–March. (c) ACC(w) and (d) MSE(w) for weekly frequency of CESM2 11-member ensemble mean WRs centered on the calendar day (i.e., [d − 3, d + 3] for a lead of d days). Skill is evaluated relative to ERA5 WRs. Shading indicates the (two tailed) 99% confidence level using a 10 000-member bootstrap. WRs are indicated in the legend.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

We can also assess the agreement between CESM2 and ERA5 weather regimes by calculating ACC and MSE on a rolling 7-day basis, which reduces error penalization if a weather regime forecast is off by a day or so (Robertson et al. 2020). In this case, ACC and MSE are computed as
ACC(w)=j=1n(w)(ojo¯)(fjf¯)j=1n(w)(ojo¯)2j=1n(w)(fjf¯)2 and
MSE(w)=1n(w)j=1n(w)(ojfj)2,
where n(w) is the number of samples within a centered 7-day sliding window w, where w includes all samples within d − 3, d − 2, d − 1, d, d + 1, d + 2, and d + 3 (for d = [4, 39], where d is lead day). Given the comparatively larger sample size, we are able to compute ACC(w) and MSE(w) stratified by weather regime, where oj and fj are the observations and forecasts (respectively) expressed as frequency of a specific weather regime class within the centered 7-day sliding window w. ACC(w) and MSE(w) are shown in Figs. 7c and 7d.

ACC(w) shows very small differences across weather regimes as a function of lead time (Fig. 7c). MSE(w) shows more persistent results, with the West Coast high regime having a higher MSE than all other regimes across all lead times (Fig. 7d). Since the West Coast high regime has a positive bias (overoccurrence) during later lead times, and MSE gives more weight to larger differences by definition, the higher MSE(w) result is expected. The Alaskan ridge and Greenland high regimes consistently show lower MSE(w) at lead times within the subseasonal time scale whereas the Pacific trough regime consistently has the second highest MSE(w) (Fig. 7), all of which are in agreement with their respective frequency biases.

d. Ensemble spread of weather regimes

We next assess how often the same weather regime is predicted by members of the CESM2 ensemble at a given lead time. We refer to ensemble member consistency as “ensemble agreement.” Fig. 8 shows the frequency (left y axis) of ensemble agreement as a function of lead time (x axis). A subjective threshold is set for the extent of ensemble agreement considered as a function of lead time, which is shown with the dotted line in Fig. 8 (right y axis). A threshold of 11 ensemble members is considered during early lead times and a lower threshold of 7 ensemble members is considered during later lead times because ensemble agreement will decrease due to ensemble dispersion. The frequency at which the ERA5 “observed” weather regime was consistent with the CESM2 predicted weather regime is also shown with black bars in Fig. 8 for comparison.

Fig. 8.
Fig. 8.

Frequency of CESM2 11-member ensemble agreement exceedance (following the left-side y axis) for a subjectively set threshold (as shown with the dashed line following the right-side y axis) that gradually decreases as lead time increases (days, shown on the x axis) for (a) WR1, (b) WR2, (c) WR3, and (d) WR4. CESM2 is shown in blue, and ERA5 is shown in black, as indicated in the legend in (a). The ERA5 black bars represent the frequency at which the ERA5 observed WR was consistent with the CESM2 predicted WR and is shown for all lead times (thus, the black bar will never exceed the blue bar). (left) The first 10 lead days are shown to better visualize the higher frequencies.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Within the first 10 days of the initialized CESM2 reforecasts, there is substantially higher ensemble agreement for the West Coast high regime as compared with other weather regimes (Fig. 8). Focusing on short lead times, there are consistently less than 60 reforecasts that have all 11 ensemble members agree on the Alaskan ridge regime (Fig. 8d), whereas over 70 reforecasts have 11 ensemble members agree on the West Coast high regime (Fig. 8a). Overall, there is generally good agreement between ERA5 and CESM2 during these early lead times and thus high skill for weather regime predictions with high ensemble agreement. We note that it is possible that initialization shock may be playing a role during these early lead times and influencing the increased ensemble member spread associated with the Alaskan ridge regime.

By lead days 6–9, the number of reforecasts where at least 10 ensemble members agree on a specific weather regime decreases, with the West Coast high and Pacific trough regimes at about 20–30 events and the Greenland high and Alaskan ridge regimes at about 15 events (Fig. 8). For lead day 20 and beyond, the West Coast high and Pacific trough regimes continue to exhibit higher ensemble agreement (with at least 7 ensemble members) as compared with the Greenland high and Alaskan ridge regimes (Fig. 8). While reforecast skill generally decreases during later lead times (fewer CESM2 reforecasts coincide with ERA5 for day 20 and beyond), the West Coast high and Pacific trough regimes exhibit comparatively higher skill (Fig. 8). In fact, there are several later lead times where there is no agreement between CESM2 and ERA5 for the Greenland high and Alaskan ridge regimes (Figs. 8c,d). Thus, the West Coast high and Pacific trough regimes exhibit higher predictability than other regimes (Figs. 8a,b). The Greenland high and Alaskan ridge regimes are comparatively less predictable and later lead times exhibit ensemble underdispersion. We note that there are numerous cases in which high ensemble agreement did not translate into skillful forecasts for the West Coast high and Pacific trough regimes (Figs. 8a,b).

e. Case studies: Precursors for forecasts of opportunity

Several reforecasts underwent what we term herein “ensemble realignment,” where ensemble trajectories that dispersed as lead time increased ended up converging again during a later lead time. This ensemble realignment, defined as at least 9 of 11 ensemble members agreeing on a specific weather regime forecast after some dispersion before week 2, occurred with more than 60 unique CESM2 reforecasts during the 1999–2019 extended cool season (of 543 reforecasts in total). Here we focus on a select few for brevity. A list of CESM2 reforecasts that realigned to have at least 9 of 11 ensemble members agree on a correctly predicted weather regime during a lead time of 2 weeks or greater is available in Table S2 in the online supplemental material.

CESM2 initialized using data from 10 February 2003 incorrectly predicted the weather regime by lead day 9, with only 5 ensemble members converging on this solution (Fig. 9a). By lead day 16, however, 10 ensemble members correctly predicted the West Coast high regime (Fig. 9a). An increase in skill during this later period can also be observed with outgoing longwave radiation (OLR) upstream of North America; ACC increased from approximately −0.5 during lead day 5 to +0.7 by lead day 16 over the North Pacific (Fig. 9b). OLR skill across the Indian–Pacific Ocean remained generally skillful throughout the full forecast period, where the Madden–Julian oscillation (MJO; Madden 1986) is typically manifested (Fig. 9b). During forecast initialization, the MJO was in phase 1, according to the definition of Wheeler and Hendon (2004), albeit very weak with an amplitude of 0.6 on 10 February 2003. El Niño was active; the oceanic Niño index (ONI) measured 0.6 during January–March (JFM). Skillful prediction of the West Coast high regime also resulted in positive ACC over North America (Figs. 9c–f) for temperature (+0.52) and precipitation (+0.33), which are higher ACC values than for the full CESM2 reforecast set (Richter et al. 2022). We note limitations in amplitude, with anomalies that are not cold enough across the southern plains and not dry enough across the Midwest (Figs. 9e,f). The positive precipitation anomalies were also seemingly missed across the desert Southwest (Fig. 9f).

Fig. 9.
Fig. 9.

CESM2 reforecasts initialized 10 Feb 2003: (a) CESM2 11-member ensemble agreement as a function of lead (days), with markers color coded on the basis of the respective WR and data product used as indicated in the legend. (b) ACC for OLR as a function of lead (days) across the Indian–Pacific Ocean (10°S–10°N, 30°E–150°W) and the North Pacific (10°–48°N, 170°E–110°W) across ocean and land areas. (middle) NOAA CPC and (bottom) CESM2 11-ensemble-member mean (c),(e) temperature and (d),(f) precipitation anomalies for lead days of highest ensemble agreement [indicated in the color-bar label and (a)]. The respective ACC and MSE are shown in (e) and (f) computed between NOAA CPC and CESM2 over land areas only.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

The CESM2 reforecast initialized with 9 November 2015 data incorrectly predicted the weather regime by lead day 10 with low confidence (4 ensemble members; Fig. 10a). By lead day 30, however, 9 ensemble members correctly predicted the Pacific trough regime. Similar to the previous case study, skill increase during the ensemble realignment period can also be observed in OLR over the North Pacific; ACC increased from approximately −0.5 during lead day 11 to +0.8 by lead day 23 (Fig. 10b). The MJO was in phase 4 (1.9 amplitude) on 9 November 2015, characterized by enhanced convection over the Maritime Continent that can result in a subsequent northward shift of the jet stream over North America [as detailed in Becker et al. (2011)]. A strong second-year El Niño was also underway (ONI of 2.6 during OND). We note that Vigaud et al. (2018) found a correlation between MJO phases 4–6, El Niño, and the Pacific trough weather regime. Skillful prediction with high certainty also resulted in positive ACC over North America at +0.77 for temperature and +0.28 for precipitation (Figs. 10c–f), both higher than the full CESM2 reforecast set (Richter et al. 2022). We again note limitations in anomaly magnitudes, with anomalies that are not warm enough across the Canadian prairies and not dry enough across the Midwest and Gulf Coast (Figs. 10e,f).

Fig. 10.
Fig. 10.

As in Fig. 9, but for CESM2 reforecasts initialized 9 Nov 2015.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

f. Case studies: Precursors limiting predictability

Here we explore cases where ensemble realignment occurred during later lead times but did not result in a skillful forecast. CESM2 initialized using data from 29 December 2014 had 10 ensemble members incorrectly predict the Pacific trough regime during lead days 18 and 19 (Fig. 11a). MJO was in phase 4 (amplitude of 1.0) on 29 December 2014 and El Niño was active with an ONI of 0.7 during NDJ. During the same period, OLR exhibited poor ACC across both the Indian–Pacific Ocean and the North Pacific (Fig. 11b). The poor skill is also reflected in the temperature and precipitation anomalies during the same period across North America with ACC of +0.39 and −0.01, respectively (Figs. 11c–f). Large amplitude errors are also evident, which are reflected in the large MSE value for 2-m temperature. This case study provides further evidence that tropical–extratropical teleconnections contribute to North American weather patterns and that the atmosphere over the North Pacific serves an important role as a causal pathway. Causal pathways are the hypothesized direct and indirect linkages between variables and outcomes (Kretschmer et al. 2021).

Fig. 11.
Fig. 11.

As in Fig. 9, but for CESM2 reforecasts initialized 29 Dec 2014.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Another example of substantial ensemble realignment resulting in an incorrect forecast is the CESM2 11 member ensemble initialized with data from 5 December 2016 (Fig. 12). During lead days 32 and 33, 10 ensemble members incorrectly predicted the Pacific trough regime (the weather regime observed during that time was the Alaskan ridge; Fig. 12a). Unlike the previous example (Fig. 11), OLR fields were represented skillfully, with ACC at approximately +0.6 across the Indian–Pacific Ocean and +0.5 across the North Pacific (Fig. 12b). At initialization, the MJO was in phase 2 albeit quite weak (amplitude of 0.2) and La Niña was active with an ONI of −0.6 during NDJ. ACC for 2-m temperature and precipitation are −0.39 and +0.22 and substantial amplitude biases are also reflected in large MSE values. Thus, while skillful representation of OLR across areas upstream of North America can contribute to skillful prediction of large-scale weather regimes and their associated temperature and precipitation anomalies, model systematic biases (underoccurrence of the Alaskan ridge regime; Fig. 3) can interfere during later lead times and result in incorrect and high-confidence predictions.

Fig. 12.
Fig. 12.

As in Fig. 9, but for CESM2 reforecasts initialized 5 Dec 2016.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

4. Conclusions

In this study, we assessed the climatological representation and predictability of North American weather regimes within the CESM2 initialized subseasonal prediction system (Richter et al. 2022), which is a new research prediction system contributing to the SubX suite of models (Pegion et al. 2019). Four key North American weather regimes were identified: (i) West Coast high, (ii) Pacific trough, (iii) Greenland high, and (iv) Alaskan ridge. Overall, CESM2 exhibits similar weather regime frequencies to those identified within ERA5 (see section 3a). We also found that 2-m temperature and precipitation patterns associated with weather regimes are distinct across North America within CESM2, albeit with amplitude biases (lower-magnitude anomalies during later lead times as compared with ERA5), potentially related to ensemble spread or systematic errors within CESM2 as the forecasts drift from the initial state (see section 3b). The West Coast high and Pacific trough regimes exhibited higher predictability than the Greenland high and Alaskan ridge regimes, evident by more skillful forecasts with comparatively lower ensemble spread (see section 3d). The enhanced predictability of the West Coast high and Pacific trough regimes is partly related to El Niño, the latter of which is consistent with Vigaud et al. (2018) (see appendix B). Skillful prediction of all weather regimes could be partly linked to skillful representation of OLR and SSTs across the tropical and extratropical North Pacific, which can serve as a causal pathway between the tropics and North America and are strongly influenced by El Niño–Southern Oscillation (ENSO) and the MJO (e.g., Zhang 2013).

Chaos theory is based on the principle that small errors in initial conditions or numerical model errors (largely due to physical system approximations) will result in initialized forecasts drifting away from the true solution over time (Buizza 2002). The initialization of numerous ensemble members can thus help to generate an uncertainty distribution for forecasts as ensemble members will inevitably disperse as lead time increases. Within CESM2, ensemble member dispersion does indeed generally increase with increasing lead time. However, we also found numerous initialized CESM2 reforecasts that had ensemble members disperse only to then largely agree again on a prediction at a lead time of 2 weeks or greater. An occasional decrease in ensemble spread during later lead times was also recently observed in monthly-to-seasonal initialized forecasts created with the NASA Goddard Earth Observing System (Massoud et al. 2023), suggesting that this phenomenon may extend to other time scales and modeling systems. Analysis of several “ensemble realignment” case studies suggested that skillful representation of OLR over the North Pacific is important for skillful forecasts with high certainty (see section 3e). OLR across the North Pacific is related to ENSO and the MJO; ENSO and/or MJO-related diabatic heating from convective activity over the tropics induces Rossby waves that migrate poleward and eastward and influence the extratropical atmospheric circulation (Lin and Brunet 2018). However, systematic biases within the CESM2 initialized prediction system, such as overoccurrence of the West Coast high regime and underoccurrence of the Alaskan ridge and Greenland high regimes during later lead times, can interfere with skillful predictions despite skillful representation of fields within the pathways between the tropics and extratropics (see section 3f).

Continued advancements in subseasonal prediction remains of utmost importance given that societal benefits to be gained are numerous (White et al. 2017). However, skillful weather regime prediction will not benefit everyone across North America. While certain weather regimes produce robust temperature and precipitation anomalies across certain areas at the subseasonal time scale within CESM2, such as the West Coast of the United States during the Alaskan ridge regime, other areas have near-zero anomalies and/or are not statistically significant, and thus skillful prediction of weather regimes may not be particularly useful for these communities. We also note that while SSTs largely remained skillful throughout the forecast window considered in this study (likely due in part to oceanic memory), there are discrepancies along some coastal areas that could potentially result in upstream moisture flux modulations that may contribute to a localized reduction of precipitation skill due to upstream air–sea interactions (e.g., Molina and Allen 2020).

Future work will explore extending and improving prediction skill of weather regimes and focus on better understanding sources and limits of predictability by exploiting machine learning and explainable artificial intelligence (AI; McGovern et al. 2019; Toms et al. 2020). While we found that skillful representation of OLR patterns across the extratropics upstream of North America can contribute to skillful and highly confident predictions during later lead times, future work should further quantify causal pathways associated with weather regimes (Kretschmer et al. 2021). Further exploration of the interplay between ENSO and the MJO on weather regime predictability should also be pursued (Arcodia et al. 2020). More broadly, our study highlights how unsupervised learning can be leveraged for pattern discovery across subseasonal time scales, including uncovering sources and limits of predictability within initialized prediction systems.

1

From a total sample size of 23 392.

Acknowledgments.

This material is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of Biological and Environmental Research (BER), Regional and Global Model Analysis (RGMA) component of the Earth and Environmental System Modeling Program under Award Number DE-SC0022070 and National Science Foundation (NSF) IA 1947282. This work was also supported by the National Center for Atmospheric Research (NCAR), which is a major facility sponsored by the NSF under Cooperative Agreement 1852977. The CESM project is supported primarily by the National Science Foundation (NSF). Computing and data storage resources on Cheyenne and Casper were provided by the Computational and Information Systems Laboratory (CISL) at NCAR.

Data availability statement.

CESM2 (CAM6) subseasonal reforecast outputs are available for download from the NCAR Climate Data Gateway (https://doi.org/10.5065/0s63-m767). ERA5 data used in this study can be obtained from the NCAR Research Data Archive (https://doi.org/10.5065/D6X34W69). MJO observations are available from the Australian Bureau of Meteorology (http://www.bom.gov.au/climate/mjo/). The oceanic Niño index is available from NOAA CPC (https://origin.cpc.ncep.noaa.gov/products/analysis_monitoring/ensostuff/ONI_v5.php). Software developed for this study is available as open source at the GitHub repository (https://github.com/mariajmolina/ML-for-S2S).

APPENDIX A

Additional Methodology Details

We address lead-time bias correction. As described in Pegion et al. (2019), the climatology used to calculate the anomalies is created by first taking the ensemble mean for individual days of each forecast and then taking a multiyear mean of forecasts for each day of the year. Since CESM2 was initialized every week on Mondays, this results in varying sample sizes for a given calendar day. To ameliorate resultant noise and data sparseness for certain calendar days, a 31-day triangular sliding window (centered with ±15 days) is applied periodically along the time dimension for each lead time, which can result in a climatology for a given calendar day that differs based on the respective lead time. Such differences based on lead time are expected because as lead time increases, the model state will generally move from initial conditions to its model-intrinsic state. See Pegion et al. (2019) for evidence that this method is consistent with fitting harmonics to data (Tippett et al. 2018).

APPENDIX B

Additional Analysis on Conditions Upstream of Weather Regimes

OLR is often used as a proxy for convection and precipitation (Matsumoto and Murakami 2000; Sandeep and Stordal 2013), where negative OLR anomalies are associated with reduced thermal radiation emitted to space, which is representative of convection and wetter conditions. OLR across the tropical Pacific can also be associated with ENSO (Wang et al. 2017; Capotondi et al. 2020; Molina et al. 2022). For example, El Niño results in warm SST anomalies across the eastern equatorial Pacific and can contribute to enhanced overlying convection (negative OLR anomalies). Large-scale OLR and SST patterns upstream of North America across the subtropics and tropics can shed light on precursor mechanisms associated with weather regimes that can result in anomalous temperature, precipitation, and extremes (e.g., Alexander et al. 2002; Molina et al. 2018; Jong et al. 2020).

Here we focus on upstream precursor patterns within CESM2, with weeks 1–2 OLR and SSTs as precursors for weeks 3–4 weather regimes (Fig. B1), weeks 3–4 OLR and SSTs as precursors for weeks 5–6 weather regimes (Fig. B2), and weeks 1–2 OLR and SSTs as precursors for weeks 5–6 weather regimes (Fig. B3). Figures B1B3 were created as follows: (i) with ACC as a metric, the top 25% OLR or SST fields upstream of North America were extracted from all upstream data, (ii) the remaining downstream fields were stratified into their respective weather regimes, and (iii) the ACC for downstream 500-hPa geopotential height fields were computed (both top 25% and overall ACC for each weather regime are shown).

Fig. B1.
Fig. B1.

CESM2 composites of the top 25% ACC (using ERA5 as observations) weeks 1–2 mean daily OLR (W m−2) and SST (°C) anomalies stratified by weeks 3–4 WR: CESM2 (color bars) and ERA5 data (black contour lines: the dashed line shows −5 and the solid line shows +5 for OLR in the left column, the dashed line shows −2.5 and the solid line shows +2.5 for OLR in the center column, and the dashed line shows −0.5 and the solid line shows +0.5 for SSTs in the right column) are shown for (a),(d),(g),(j) the tropical Pacific (10°S–10°N, 30°E–150°W) OLR; (b),(e),(h),(k) North Pacific (10°N–48°N, 170°E–110°°W) OLR; and (c),(f),(i),(l) tropical Pacific (10°S–10°N, 120°E–80°W) SSTs. Corresponding ACCs for weeks 3–4 WRs are shown within each plot, computed using CESM2 and ERA5 500-hPa geopotential height anomalies over North America.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Fig. B2.
Fig. B2.

As in Fig. B1, but for weeks 3–4 OLR and SSTs, and weeks 5–6 North American WRs.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Fig. B3.
Fig. B3.

As in Fig. B1, but for weeks 1–2 OLR and SSTs, and weeks 5–6 North American WRs.

Citation: Artificial Intelligence for the Earth Systems 2, 2; 10.1175/AIES-D-22-0051.1

Figures show that skillful representation of upstream precursor patterns during weeks 1–2 within CESM2 can contribute to skillful prediction of weeks 3–4 weather regimes over North America. For example, weeks 1–2 OLR patterns across the Indo-Pacific and western Pacific Ocean that score within the top 25% ACC are associated with more skillful weeks 3–4 500-hPa geopotential height patterns over North America, specifically for the Greenland high regime (Fig. B1g). OLR anomalies over the tropical Pacific show convection over the west tropical Pacific, which would coincide with MJO phase 6 and is consistent with results from Vigaud et al. (2018) for the weather regime characterized by anomalous high pressure over the Arctic therein. Skillfully represented (within the top 25% ACC) weeks 1–2 OLR patterns across the North Pacific are also associated with higher skill in weeks 3–4 500-hPa geopotential height anomalies over North America for the Alaskan ridge regime (Fig. B1k). Some higher skill is also evident when considering skillful representation of weeks 1–2 OLR patterns across the North Pacific for the Greenland high regime during weeks 3–4 (Fig. B1h).

The Alaskan ridge regime appears to be more frequently associated with a La Niña-like SST signal across the equatorial Pacific, whereas the Pacific trough regime appears to be more frequently associated with an El Niño–like SST signal (Figs. B1l,f), consistent with Vigaud et al. (2018). As described in Vigaud et al. (2018), ENSO can strongly influence variability in the position of the jet stream over the North Pacific, resulting in downstream modulations to large-scale weather regime patterns (Athanasiadis et al. 2010; Delcambre et al. 2013). However, it is likely that tropical Pacific SSTs are not the primary source of predictability for weeks 3–4 (Mariotti et al. 2018) and that some predictability could also be arising from land or atmospheric components.

As lead time increases, skillful representation (within the top 25% ACC) of upstream OLR and SST patterns during weeks 3–4 is associated with higher skill in weeks 5–6 500-hPa geopotential height patterns over North America during the Pacific trough regime (Fig. B2d). Similarly, weeks 3–4 OLR patterns across the North Pacific and SST patterns across the equatorial Pacific that score within the top 25% ACC are associated with more skillful weeks 5–6 500-hPa geopotential height patterns over North America during the Pacific trough regime (Figs. B2e,f). Interestingly, the more skillful representation of upstream OLR and SST patterns during weeks 3–4 are not associated with more skillful weeks 5–6 500-hPa geopotential height anomalies across North America during the Alaskan ridge regime (Figs. B2j–l).

Skillful representation of OLR and SST patterns during earlier lead times can also potentially provide predictive information for later lead times (e.g., weeks 5–6). Most of the predictive skill for downstream weeks 5–6 500-hPa geopotential height anomalies over North America during the Pacific trough regime seems to be at least partly associated with skillful representation (within the top 25% ACC) of weeks 1–2 OLR and SST patterns (Figs. B3d–f). Associated skill gained with skillful representation of upstream OLR or SST patterns during weeks 1–2 for weeks 5–6 seems to be marginal for other weather regimes (Fig. B3).

REFERENCES

  • Adler, R. F., and Coauthors, 2003: The Version-2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979–present). J. Hydrometeor., 4, 11471167, https://doi.org/10.1175/1525-7541(2003)004<1147:TVGPCP>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Albers, J. R., and M. Newman, 2019: A priori identification of skillful extratropical subseasonal forecasts. Geophys. Res. Lett., 46, 12 52712 536, https://doi.org/10.1029/2019GL085270.

    • Search Google Scholar
    • Export Citation
  • Alexander, M. A., I. Bladé, M. Newman, J. R. Lanzante, N.-C. Lau, and J. D. Scott, 2002: The atmospheric bridge: The influence of ENSO teleconnections on air–sea interaction over the global oceans. J. Climate, 15, 22052231, https://doi.org/10.1175/1520-0442(2002)015<2205:TABTIO>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Amini, S., and D. M. Straus, 2019: Control of storminess over the Pacific and North America by circulation regimes. Climate Dyn., 52, 47494770, https://doi.org/10.1007/s00382-018-4409-7.

    • Search Google Scholar
    • Export Citation
  • Arcodia, M. C., B. P. Kirtman, and L. S. P. Siqueira, 2020: How MJO teleconnections and ENSO interference impacts U.S. precipitation. J. Climate, 33, 46214640, https://doi.org/10.1175/JCLI-D-19-0448.1.

    • Search Google Scholar
    • Export Citation
  • Arthur, D., and S. Vassilvitskii, 2006: k-means++: The advantages of careful seeding. Stanford University Tech. Rep., 9 pp., https://theory.stanford.edu/∼sergei/papers/kMeansPP-soda.pdf.

  • Athanasiadis, P. J., J. M. Wallace, and J. J. Wettstein, 2010: Patterns of wintertime jet stream variability and their relation to the storm tracks. J. Atmos. Sci., 67, 13611381, https://doi.org/10.1175/2009JAS3270.1.

    • Search Google Scholar
    • Export Citation
  • Barnes, E. A., K. Mayer, B. Toms, Z. Martin, and E. Gordon, 2020: Identifying opportunities for skillful weather prediction with interpretable neural networks. arXiv, 2012.07830v1, https://doi.org/10.48550/arXiv.2012.07830.

  • Barthlott, C., and C. Hoose, 2018: Aerosol effects on clouds and precipitation over central Europe in different weather regimes. J. Atmos. Sci., 75, 42474264, https://doi.org/10.1175/JAS-D-18-0110.1.

    • Search Google Scholar
    • Export Citation
  • Bauer, P., A. Thorpe, and G. Brunet, 2015: The quiet revolution of numerical weather prediction. Nature, 525, 4755, https://doi.org/10.1038/nature14956.

    • Search Google Scholar
    • Export Citation
  • Becker, E. J., E. H. Berbery, and R. W. Higgins, 2011: Modulation of cold-season U.S. daily precipitation by the Madden–Julian oscillation. J. Climate, 24, 51575166, https://doi.org/10.1175/2011JCLI4018.1.

    • Search Google Scholar
    • Export Citation
  • Berner, J., and G. Branstator, 2007: Linear and nonlinear signatures in the planetary wave dynamics of an AGCM: Probability density functions. J. Atmos. Sci., 64, 117136, https://doi.org/10.1175/JAS3822.1.

    • Search Google Scholar
    • Export Citation
  • Bruyère, C., C. Raktham, J. Done, J. Kreasuwun, C. Thongbai, and W. Promnopas, 2017: Major weather regime changes over Southeast Asia in a near-term future scenario. Climate Res., 72 (1), 118, https://doi.org/10.3354/cr01442.

    • Search Google Scholar
    • Export Citation
  • Büeler, D., L. Ferranti, L. Magnusson, J. F. Quinting, and C. M. Grams, 2021: Year-round sub-seasonal forecast skill for Atlantic–European weather regimes. Quart. J. Roy. Meteor. Soc., 147, 42834309, https://doi.org/10.1002/qj.4178.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., 2002: Chaos and weather prediction January 2000. European Centre for Medium-Range Weather Forecasting Meteorological Training Course Lecture Series, 28 pp., https://www.ecmwf.int/sites/default/files/elibrary/2002/16927-chaos-and-weather-prediction.pdf.

  • Buizza, R., P. L. Houtekamer, G. Pellerin, Z. Toth, Y. Zhu, and M. Wei, 2005: A comparison of the ECMWF, MSC, and NCEP global ensemble prediction systems. Mon. Wea. Rev., 133, 10761097, https://doi.org/10.1175/MWR2905.1.

    • Search Google Scholar
    • Export Citation
  • Capotondi, A., C. Deser, A. S. Phillips, Y. Okumura, and S. M. Larson, 2020: ENSO and Pacific decadal variability in the Community Earth System Model version 2. J. Adv. Model. Earth Syst., 12, e2019MS002022, https://doi.org/10.1029/2019MS002022.

    • Search Google Scholar
    • Export Citation
  • Cassou, C., 2008: Intraseasonal interaction between the Madden–Julian Oscillation and the North Atlantic Oscillation. Nature, 455, 523527, https://doi.org/10.1038/nature07286.

    • Search Google Scholar
    • Export Citation
  • Celebi, M. E., and K. Aydin, 2016: Unsupervised Learning Algorithms. Springer, 558 pp., https://doi.org/10.1007/978-3-319-24211-8.

  • Chernick, M. R., 2011: Bootstrap Methods: A Guide for Practitioners and Researchers. Wiley Series in Probability and Statistics, John Wiley and Sons, 400 pp.

  • Christiansen, B., 2007: Atmospheric circulation regimes: Can cluster analysis provide the number? J. Climate, 20, 22292250, https://doi.org/10.1175/JCLI4107.1.

    • Search Google Scholar
    • Export Citation
  • Conway, D., and P. Jones, 1998: The use of weather types and air flow indices for GCM downscaling. J. Hydrol., 212213, 348361, https://doi.org/10.1016/S0022-1694(98)00216-9.

    • Search Google Scholar
    • Export Citation
  • Cortesi, N., V. Torralba, L. Lledó, A. Manrique-Suñén, N. Gonzalez-Reviriego, A. Soret, and F. J. Doblas-Reyes, 2021: Yearly evolution of Euro-Atlantic weather regimes and of their sub-seasonal predictability. Climate Dyn., 56, 39333964, https://doi.org/10.1007/s00382-021-05679-y.

    • Search Google Scholar
    • Export Citation
  • Danabasoglu, G., S. C. Bates, B. P. Briegleb, S. R. Jayne, M. Jochum, W. G. Large, S. Peacock, and S. G. Yeager, 2012: The CCSM4 ocean component. J. Climate, 25, 13611389, https://doi.org/10.1175/JCLI-D-11-00091.1.

    • Search Google Scholar
    • Export Citation
  • Danabasoglu, G., and Coauthors, 2020: The Community Earth System Model version 2 (CESM2). J. Adv. Model. Earth Syst., 12, e2019MS001916, https://doi.org/10.1029/2019MS001916.

    • Search Google Scholar
    • Export Citation
  • Delcambre, S. C., D. J. Lorenz, D. J. Vimont, and J. E. Martin, 2013: Diagnosing Northern Hemisphere jet portrayal in 17 CMIP3 global climate models: Twenty-first-century projections. J. Climate, 26, 49304946, https://doi.org/10.1175/JCLI-D-12-00359.1.

    • Search Google Scholar
    • Export Citation
  • Deser, C., and Coauthors, 2020: Insights from Earth system model initial-condition large ensembles and future prospects. Nat. Climate Change, 10, 277286, https://doi.org/10.1038/s41558-020-0731-2.

    • Search Google Scholar
    • Export Citation
  • Elkan, C., 2003: Using the triangle inequality to accelerate k-means. Proc. 20th Int. Conf. on Machine Learning, Washington, DC, Association for the Advancement of Artificial Intelligence, 147–153.

  • Ferranti, L., S. Corti, and M. Janousek, 2015: Flow-dependent verification of the ECMWF ensemble over the Euro-Atlantic sector. Quart. J. Roy. Meteor. Soc., 141, 916924, https://doi.org/10.1002/qj.2411.

    • Search Google Scholar
    • Export Citation
  • Ferranti, L., L. Magnusson, F. Vitart, and D. S. Richardson, 2018: How far in advance can we predict changes in large-scale flow leading to severe cold conditions over Europe? Quart. J. Roy. Meteor. Soc., 144, 17881802, https://doi.org/10.1002/qj.3341.

    • Search Google Scholar
    • Export Citation
  • Harris, C. R., and Coauthors, 2020: Array programming with NumPy. Nature, 585, 357362, https://doi.org/10.1038/s41586-020-2649-2.

  • Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 19992049, https://doi.org/10.1002/qj.3803.

    • Search Google Scholar
    • Export Citation
  • Hochman, A., G. Messori, J. F. Quinting, J. G. Pinto, and C. M. Grams, 2021: Do Atlantic-European weather regimes physically exist? Geophys. Res. Lett., 48, e2021GL095574, https://doi.org/10.1029/2021GL095574.

    • Search Google Scholar
    • Export Citation
  • Hoyer, S., and J. Hamman, 2017: xarray: N-D labeled arrays and datasets in Python. J. Open Res. Software, 5, 10, https://doi.org/10.5334/jors.148.

    • Search Google Scholar
    • Export Citation
  • Huang, W. T. K., A. Charlton-Perez, R. W. Lee, R. Neal, C. Sarran, and T. Sun, 2020: Weather regimes and patterns associated with temperature-related excess mortality in the UK: A pathway to sub-seasonal risk forecasting. Environ. Res. Lett., 15, 124052, https://doi.org/10.1088/1748-9326/abcbba.

    • Search Google Scholar
    • Export Citation
  • Hunke, E. C., W. H. Lipscomb, A. K. Turner, N. Jeffery, and S. Elliott, 2015: CICE: The Los Alamos Sea Ice Model—Documentation and software user’s manual, version 5.1. Los Alamos National Laboratory Doc. LA-CC-06-012, 116 pp., http://www.ccpo.odu.edu/∼klinck/Reprints/PDF/cicedoc2015.pdf.

  • Hunter, J. D., 2007: Matplotlib: A 2D graphics environment. Comput. Sci. Eng., 9, 9095, https://doi.org/10.1109/MCSE.2007.55.

  • Hurrell, J. W., Y. Kushnir, and M. Visbeck, 2001: The North Atlantic Oscillation. Science, 291, 603605, https://doi.org/10.1126/science.1058761.

    • Search Google Scholar
    • Export Citation
  • Jong, B.-T., M. Ting, R. Seager, and W. B. Anderson, 2020: ENSO teleconnections and impacts on US summertime temperature during a multiyear La Niña life cycle. J. Climate, 33, 60096024, https://doi.org/10.1175/JCLI-D-19-0701.1.

    • Search Google Scholar
    • Export Citation
  • Kageyama, M., F. D’Andrea, G. Ramstein, P. J. Valdes, and R. Vautard, 1999: Weather regimes in past climate atmospheric general circulation model simulations. Climate Dyn., 15, 773793, https://doi.org/10.1007/s003820050315.

    • Search Google Scholar
    • Export Citation
  • Kimoto, M., and M. Ghil, 1993a: Multiple flow regimes in the Northern Hemisphere winter. Part I: Methodology and hemispheric regimes. J. Atmos. Sci., 50, 26252644, https://doi.org/10.1175/1520-0469(1993)050<2625:MFRITN>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Kimoto, M., and M. Ghil, 1993b: Multiple flow regimes in the Northern Hemisphere winter. Part II: Sectorial regimes and preferred transitions. J. Atmos. Sci., 50, 26452673, https://doi.org/10.1175/1520-0469(1993)050<2645:MFRITN>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Kretschmer, M., S. V. Adams, A. Arribas, R. Prudden, N. Robinson, E. Saggioro, and T. G. Shepherd, 2021: Quantifying causal pathways of teleconnections. Bull. Amer. Meteor. Soc., 102, E2247E2263, https://doi.org/10.1175/BAMS-D-20-0117.1.

    • Search Google Scholar
    • Export Citation
  • Lamb, H. H., 1972: British Isles Weather Types and a Register of the Daily Sequence of Circulation Patterns 1861–1971. Her Majesty’s Stationery Office, 85 pp.

  • Lawrence, D. M., and Coauthors, 2019: The Community Land Model version 5: Description of new features, benchmarking, and impact of forcing uncertainty. J. Adv. Model. Earth Syst., 11, 42454287, https://doi.org/10.1029/2018MS001583.

    • Search Google Scholar
    • Export Citation
  • Lee, S. H., J. C. Furtado, and A. J. Charlton-Perez, 2019: Wintertime North American weather regimes and the Arctic stratospheric polar vortex. Geophys. Res. Lett., 46, 14 89214 900, https://doi.org/10.1029/2019GL085592.

    • Search Google Scholar
    • Export Citation
  • Li, W., L. Li, M. Ting, and Y. Liu, 2012: Intensification of Northern Hemisphere subtropical highs in a warming climate. Nat. Geosci., 5, 830834, https://doi.org/10.1038/ngeo1590.

    • Search Google Scholar
    • Export Citation
  • Lin, H., and G. Brunet, 2018: Extratropical response to the MJO: Nonlinearity and sensitivity to the initial state. J. Atmos. Sci., 75, 219234, https://doi.org/10.1175/JAS-D-17-0189.1.

    • Search Google Scholar
    • Export Citation
  • Lorenz, E. N., 1963: Deterministic nonperiodic flow. J. Atmos. Sci., 20, 130141, https://doi.org/10.1175/1520-0469(1963)020<0130:DNF>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • MacQueen, J., 1967: Some methods for classification and analysis of multivariate observations. Fifth Berkeley Symp. on Mathematical Statistics and Probability, Vol. 1, Los Angeles, CA, University of California, 281–297.

  • Madden, R. A., 1986: Seasonal variations of the 40–50 day oscillation in the tropics. J. Atmos. Sci., 43, 31383158, https://doi.org/10.1175/1520-0469(1986)043<3138:SVOTDO>2.0.CO;2.

    • Search Google Scholar
    • Export Citation
  • Magnusson, L., J. Nycander, and E. Källén, 2009: Flow-dependent versus flow-independent initial perturbations for ensemble prediction. Tellus, 61A, 194209, https://doi.org/10.1111/j.1600-0870.2008.00385.x.

    • Search Google Scholar
    • Export Citation
  • Maher, N., S. Milinski, and R. Ludwig, 2021: Large ensemble climate model simulations: Introduction, overview, and future prospects for utilising multiple types of large ensemble. Earth Syst. Dyn., 12, 401418, https://doi.org/10.5194/esd-12-401-2021.

    • Search Google Scholar
    • Export Citation
  • Mariotti, A., P. M. Ruti, and M. Rixen, 2018: Progress in subseasonal to seasonal prediction through a joint weather and climate community effort. npj Climate Atmos. Sci., 1, 4, https://doi.org/10.1038/s41612-018-0014-z.

    • Search Google Scholar
    • Export Citation
  • Mariotti, A., and Coauthors, 2020: Windows of opportunity for skillful forecasts subseasonal to seasonal and beyond. Bull. Amer. Meteor. Soc., 101, E608E625, https://doi.org/10.1175/BAMS-D-18-0326.1.

    • Search Google Scholar
    • Export Citation
  • Massoud, E. C., L. Andrews, R. Reichle, A. Molod, J. Park, S. Ruehr, and M. Girotto, 2023: Seasonal forecasting skill for the high mountain Asia region in the Goddard Earth Observing System. Earth Syst. Dyn., 14, 147–171, https://10.5194/esd-14-147-2023.

    • Search Google Scholar
    • Export Citation
  • Matsumoto, J., and T. Murakami, 2000: Annual changes of tropical convective activities as revealed from equatorially symmetric OLR data. J. Meteor. Soc. Japan, 78, 543561, https://doi.org/10.2151/jmsj1965.78.5_543.

    • Search Google Scholar
    • Export Citation
  • Mayer, K. J., and E. A. Barnes, 2021: Subseasonal forecasts of opportunity identified by an explainable neural network. Geophys. Res. Lett., 48, e2020GL092092, https://doi.org/10.1029/2020GL092092.

    • Search Google Scholar
    • Export Citation