1 Introduction

Climate change is one of the greatest challenges faced by humankind as it poses an existential threat to many aspects of the current social–ecological landscape of natural and human systems. Over North and Central America and the Caribbean, observations provide irrefutable evidence that climate change is already happening. For instance, there has been a decreasing trend in snow accumulation in the western United States (U.S.) during the cold season over the last several decades (Mote et al. 2005; Ashfaq et al. 2013). Similarly, several regions are experiencing increasing trends in minimum temperatures (Batibeniz et al. 2020) and more intense and widespread precipitation extremes (Rastogi et al. 2020). Mean temperatures are also increasing in southwestern U.S. and northwestern Mexico (Cavazos et al. 2020). An anthropogenic footprint is also detectable in the greater frequency of multivariate extremes, such as droughts (Diffenbaugh et al. 2015) and wildfires (Radeloff et al. 2018; Abatzoglou and Williams 2016). Studies suggest that continued greenhouse gas emissions will lead to further warming, along with long-term changes in all components of the climate system, increasing the likelihood of widespread and potentially irreversible impacts on social and ecological systems (IPCC 2014; Sillmann et al. 2013; Pu et al. 2020; Christensen et al. 2013; Fischer and Knutti 2016; Easterling et al. 2017; Diaconescu et al. 2018; Giorgi et al. 2018; Moreno et al. 2019; Veas-Ayala et al. 2018; Oppenheimer et al. 2019; Ashfaq et al. 2020; Cook et al. 2020; Wang et al. 2020a, b). These expected changes in climate characteristics will likely intensify the regional hydrological cycle (Ashfaq et al. 2013; Naz et al. 2016), exacerbate the risk of water stress in the regions where river runoff is dominated by snowmelt (Pagan et al. 2016), and increase the population’s exposure to climate extremes (Batibeniz et al. 2020).

General circulation models (GCMs) are the most sophisticated tools available to investigate the climate system response to increases in radiative forcing and to identify the mechanisms driving that response (Taylor et al. 2012; Flato et al. 2013). However, despite potential progress over the last two decades in GCM simulations of past, present, and future climate, considerable regional biases, and other inadequacies still exist due to incomplete representation of key regional scale processes, poor parameterizations, imperfect initial conditions, and coarse resolutions (van Der Wiel et al. 2016; Wehner et al. 2014; Diallo et al. 2019). These modeling deficiencies are in addition to the influences of internal variability of the Earth system that collectively lead to uncertainty in projections of climate change at global to regional scales. To address these challenges, a coordinated framework by the climate modeling community has produced different phases of the Coupled Model Intercomparison Project (CMIP), where state-of-the-art GCMs are run with a common setup. CMIP aims to better understand past, present, and future climate change arising from both natural and forced variability, and in response to changes in radiative forcing, in a multimodel ensemble context (Eyring et al. 2016). CMIP also supports the Intergovernmental Panel on Climate Change (IPCC) Assessment Reports (AR). However, confidence in the projections of GCMs-based future climate partially depends on their skill in simulating current and past climates. Previous comparative evaluations of climate simulations have shown that CMIP Phase 5 (CMIP5) models performed better than CMIP Phase 3 (CMIP3) models, particularly over North and Central America including Caribbean (Bukovsky et al. 2015; Maloney et al. 2014; Ryu and Hayhoe 2014; Koutroulis et al. 2016; Hidalgo and Alfaro 2015).

New simulations from the latest state-of-the-art climate models participating in phase 6 of the CMIP (CMIP6) are now available (Eyring et al. 2016). These simulations provide a new opportunity to evaluate the Earth system response to change in radiative forcings during the twenty-first century. The evaluation of CMIP6 simulations over several regions, including South Asia, Africa and Arabian Peninsula, suggest that their simulated responses exhibits differences from earlier CMIPs studies (Almazroui et al. 2020a,b,c). The CMIP6 models are typically enhanced versions of the models that participated in earlier phases of CMIP. Most have improved parameterizations of cloud microphysics and better representations of various Earth system processes, such as biogeochemical cycles and ice sheets. The average resolution of CMIP6 GCMs is also finer than that of CMIP5 GCMs (Eyring et al. 2016). Several studies have evaluated CMIP6 GCM over the U.S. in the historical period. For instance, Wehner et al. (2020) reported no discernible differences between CMIP5 and CMIP6 GCMs in the simulation of daily precipitation and temperature extremes in the historical period. Similarly, Srivastava et al. (2020) evaluated daily characteristics of precipitation in the historical period and concluded that most CMIP6 models overestimate the occurrence and variability of wet spell durations over the western U.S. while they underestimate the occurrence of dry spells over most of the southern U.S. Likewise, Akinsanola et al. (2020) noted that CMIP6 models showed relatively better skill in the simulation of warm season extreme precipitation over the U.S. However, studies focused on the evaluation of future CMIP5 climatic changes over the North and Central America and the Caribbean are relatively limited and mostly thus far have focused on the North American monsoon that exhibits drying in response to increase in radiative forcing (Jin et al. 2020; He et al. 2020; Wang et al. 2020a, b). Therefore, a study assessing the CMIP6 models performance and their simulated precipitation and temperature changes in the twenty-first century over the region encompassing the contiguous U.S., Central America (including Mexico), and the Caribbean, is currently lacking.

This study addresses these gaps by assessing the future climate responses at a subregional scale over these regions through the analysis of a large suite of CMIP6 GCMs under various Shared Socioeconomic Pathways/Representative Concentration Pathways (SSP/RCP) scenarios. In addition, similar analyses are performed using a subset of GCMs that are selected based on their best performance in the historical period. The projections are divided into near-term, mid-term, and far-term time periods. The remainder of the paper is organized as follows. Section 2 describes the observed and model data, along with methods and metrics used to assess model performance and to select the best-performing models. Sections 3 and 4 discuss the results of the study for the present and future climate, respectively. A summary and conclusions are outlined in the last section.

2 Data and Methodology

This study utilized 31 CMIP6 GCMs (Table 1) (hereafter M31) that are publicly available (Eyring et al. 2016; https://esgf-node.llnl.gov/search/cmip6) to assess the accuracy of projected changes in temperature and precipitation over the contiguous U.S., Central America, and the Caribbean region (Fig. 1). For the GCM evaluation, monthly temperature observations were obtained from the Climatic Research Unit (CRU, Harris et al. 2014) and the University of Delaware (UoD, Willmott et al. 2001), and monthly precipitation observations were obtained from the CRU and the Global Precipitation Climatology Centre (GPCC, Becker et al. 2013). All CMIP6 GCM data and observations are interpolated to a common 1˚ × 1˚ latitude–longitude grid using conservative remapping (Iturbide et al. 2020).

Table 1 List of the 31 CMIP6 models used in this study along with the horizontal resolution, variant label, and
Full size table
Fig. 1
figure 1

Topography map showing the USA, Central America, and the Caribbean domain. The six subregions used in this study are: Western North America (WNA), Central North America (CNA), Eastern North America (ENA), North Central America (NCA), South Central America (SCA), and the Caribbean (CAR)

Full size image

The analysis domain was divided into six subregions: Western North America (WNA); Central North America (CNA); Eastern North America (ENA); North Central America (NCA); South Central America (SCA); and the Caribbean (CAR) (Fig. 1). The regionalization is based on that used by the Intergovernmental Panel on Climate Change (IPCC), as explained in detail by Iturbide et al. (2020). The time spans for analysis are divided into a historical reference period (1995–2014) and three future periods (near term; 2021–2040, mid term; 2041–2060, and far term; 2080–2099). The models are evaluated against observations for the reference period while future changes with respect to the reference period are calculated over four seasons, winter (December–January–February; DJF), spring (March–April–May; MAM), summer (June–July–August; JJA), and autumn (September–October–November; SON), under the low (SSP1-2.6), medium (SSP2-4.5), and high (SSP58.5) forcing scenarios, as described in Gidden et al. (2019) and O’Neill et al. (2016). The simulated future changes are considered robust if at least 66% of the models agree on the direction of change (Haensler et al. 2013; Almazroui et al. 2020a,b,c). Moreover, a two-tailed Student t test is used to test the significance of simulated changes in each grid box. In addition to spatial time–average plots, the results are also presented as time series of annual averages for each subregion. These plots show the median value of the 31 models along with the 66% likely range (i.e., 17–83 percentiles) around the median line. Trends and levels of significance are then calculated using the Mann–Kendall two-tailed test (Mann 1945; Kendall 1975; Sen 1968; Wilks 2019).

A subset of the best-performing models is also selected based on a bias analysis along with other performance measures. For simplicity, each CMIP6 model is assessed using averages calculated over the entire domain instead of over each subregion. The bias analysis is based on a threshold of 1.5 standard deviations (± 1.5 STD) about the multimodel mean bias for the historical period. According to this criterion, any model with bias exceeding the threshold is not a candidate for selection in the subset set of models (Almazroui et al. 2017a,b). Afterwards, the pattern correlation coefficient (PCC) and the root mean square error (RMSE) for annual mean temperature and precipitation are calculated. For temperature (precipitation), models with an RMSE of less than 2 °C (1 mm/day) and a pattern correlation above 0.96 (0.60) are considered for selection. Models that satisfy the selection criteria along with the RMSE and PCC thresholds for temperature and precipitation are then added to the set of selected models. The rationale for this model selection is to understand whether or not skill-based model selection has any impact on the robustness of simulated changes in temperature and precipitation in the future period (Knutti et al. 2017).

3 Analysis of Present Climate

3.1 Temperature and Rainfall Climatology

The spatial distribution of observed and simulated annual and seasonal mean temperatures averaged over the reference period 1995–2014 is shown in Fig. 2. Annual mean temperature shows a latitudinal variation ranging from 0 ºC in the north to nearly 30 ºC in the south (Fig. 2, middle and lower panel). The observed spatial and seasonal variations in temperature are well captured in the CMIP6 ensemble (Fig. 2, top row). There is substantial spatial heterogeneity in seasonal mean temperatures, particularly in winter and summer. For instance, winter temperatures fall to the deepest minima (− 12 °C) over the upper Midwest, which frequently experiences cold waves due to cold Arctic air flow, while summer average temperatures reach highs (30 °C) over Texas and northeastern Mexico as well as northwestern Mexico and Arizona. These seasonal variations in temperature patterns are well represented in the CMIP6 ensemble (Fig. 2).

Fig. 2
figure 2

Spatial distribution of annual and seasonal mean temperature (°C) obtained from the ensemble mean of M31 (upper panels), and observations (CRU: middle panels, UoD: lower panels) for the present climate, 1995–2014. The 1st, 2nd, 3rd, 4th and 5th columns represent the temperature distribution for annual, winter, spring, summer and autumn, respectively

Full size image

Figure 3 compares the distribution of precipitation between the observations and the CMIP6 ensemble for the 1995–2014 period. The simulated annual and seasonal precipitation patterns generally resemble those in the observations. However, a number of regional biases exist in seasonal precipitation magnitudes, including excess precipitation over the upper Midwest during winter and spring, a lack of spatial detail in the topography-driven winter precipitation over the western U.S., and an overly strong North American summer monsoon. Nonetheless, by and large, the simulated precipitation captures the general characteristics of the observed variations in seasonal precipitation across the study region.

Fig. 3
figure 3

Same as Fig. 2, except the spatial distribution of annual and seasonal mean precipitation (mm/month) obtained from ensemble mean of M31 (upper panels), and observations (CRU: middle panels, GPCC: lower panels) for the present climate 1995–2014

Full size image

3.2 Distribution of the Annual Cycle

The annual cycle of observed and simulated temperatures over the six subregions is shown in Fig. 4. Although there is overall consistency in the annual amplitudes between the simulations and the observations, there are some biases in the monthly magnitudes (from − 0.02 to 1.06 °C). For instance, summer (JJA) temperatures are overly warm in the CMIP6 ensemble over the WNA (1.37/1.24 °C relative to CRU/UoD), CNA (1.68/2.11), CAR (0.73/1.01), SCA (0.43/0.51), and ENA (0.55/0.82) regions, and are excessively cold over WNA (-1.50/-1.04 °C relative to CRU/UoD) in the winter (DJF) months. For the other seasons, the simulated annual cycle of temperatures in the CMIP6 ensemble is within ~ 1 °C of the observations over most of the subregions. The simulated temperature biases range from − 0.93 to 0.88 °C (− 0.02 to 1.27 °C) with respect to CRU (UoD) at annual scale.

Fig. 4
figure 4

The annual cycle of temperature (°C) obtained from observations (CRU and UoD) and the ensemble mean of M31 over six subregions during the base period (1995–2014). Only temperature over land is considered. The horizontal dashed lines represent the annual mean temperature in each dataset

Full size image

The annual cycle of observed and simulated precipitation for the six subregions is shown in Fig. 5. At annual scale, the simulated precipitation biases range from − 37.9 to 58.45% (− 33.23–51.96%) with respect to CRU (GPCC). In general, the disagreement between the observations and the CMIP6 ensemble is quite substantial in the case of precipitation. For instance, the simulated precipitation is overly strong in WNA and NCA throughout the year with particularly large biases (~ 20 mm/month) in the cold season months. In contrast, the monthly precipitation amounts in the CMIP6 ensemble are substantially lower than the observations in SCA and CAR. The simulated annual cycle exhibits both wet and dry biases over the CNA region. The dominant annual cycle over SCA, except for the central part of its Caribbean coast, is monsoonal, with highest temperatures in April–May and lowest temperatures in January (Fig. 4). Rainfall in most of SCA is characterized by two maxima in June and September, an extended dry season from November to April, and a short midsummer drought (MSD) in July–August (Fig. 5). To some extent, precipitation seasonality is explained by the seasonal strength of the North Atlantic Subtropical High (NASH) and the North–South migration of the Intertropical Convergence Zone (ITCZ) (Taylor and Alfaro 2005). The Caribbean islands have two main seasons, characterized by differences in temperature and precipitation. The wet or rainy season with higher temperatures and more accumulated precipitation occurs during the boreal summer and part of spring and autumn. The MSD is also present in most of the Caribbean, particularly in the Greater Antilles (Taylor and Alfaro, 2005). The precipitation annual cycle in Fig. 5 for SCA and CAR is in good agreement with the annual cycles previously presented over the subregions (Centella-Artola et al. 2015; Duran–Quesada et al. 2020; Maldonado et al. 2018; Martínez-Castro et al. 2018; Vichot-Llano et al. 2020a).

Fig. 5
figure 5

The annual cycle of precipitation (mm/month) obtained from observations (CRU and GPCC) and the ensemble mean of M31over six subregions for the base period 1995–2014. Only precipitation over land is considered. Note that the vertical scales of the SCA and CAR panels are different from the vertical scales of the other subregions. The horizontal dashed lines represent the mean annual precipitation in dataset

Full size image

As mentioned in the methodology section, the annual cycles shown are averages over the subregions and do not necessarily reflect the heterogeneity within the region, particularly that driven by topographic complexity over WNA, NCA, SCA, and CAR (e.g., Taylor and Alfaro 2005). However, the Caribbean MSD (Magaña et al. 1999; Amador 2008) is well reproduced by the model ensemble, even if it is underestimated relative to the observed precipitation. The MSD is also a regular feature of the southern part of NCA and the Pacific side of SCA, but it is not reflected in the area average (Curtis and Gamble 2008). This is an important limitation for stakeholders since MSD is an important aspect of the precipitation annual cycle in SCA.

3.3 Performance of Individual CMIP6 Models

Figure 6a shows the regional average of the terrestrial precipitation bias in each of the 31 CMIP6 models with respect to two sets of observations (GPCC and CRU). Most of the models have a wet bias over the analysis domain, as the simulated precipitation from 23 (21) GCMs is higher than observed while 8 (10) is lower magnitudes than observed. As a result, the ensemble mean also has a wet bias of 0.15 (0.2) mm/day as compared to GPCC (CRU). We subselect models based on their performance by using observed annual standard deviation as a threshold. Models whose bias is within the range of ± 1.5 STDs are identified as better performing models. The models whose bias exceeds ± 1.5 STD (~ ± 0.34 mm/day) are not considered for further analysis, including ACCESS–ESM1–5, CIESM, CNRM–CM6–1–HR, CNRM–CM6–1, CNRM–ESM2–1, CanESM5, IPSL–CM6A–LR, MRI–ESM2–0, and NESM3. It is interesting to note that the finest resolution GCM (CNRM–CM6–1–HR, Table 1) is not among the best-performing models for precipitation. Overall, 23 of the 31 models qualify as better performing models following this criterion.

Fig. 6
figure 6

Area-averaged bias of CMIP6 models for (a) precipitation (mm/day), and (b) temperature (°C) for the period 1995–2014. The blue and red bars in a show wet and dry bias, respectively, while in b they show cold and warm bias, respectively. The horizontal dashed and dashed-dot lines show the ensemble mean bias in M31 with respect to each of the observational datasets, as labelled. The solid lines in a and b show the magnitude of ± 1.5 standard deviation

Full size image

The biases in the simulated temperatures over the analysis domain are shown in Fig. 6b. For temperature, 19 (15) of the 31 models exhibit a warm bias when compared to the CRU (UoD) observations, resulting in a net positive bias in the ensemble mean. Among 31 models, 23 (28) simulate mean temperatures within the ± 1.5 STD range as compared to the CRU (UoD) observations. Again, the CNRM–CM6–1–HR has a large bias, as it is the coldest CMIP6 model. It should be noted that the observational datasets also differ between themselves. The ensemble mean bias is less than 0.2 °C with respect to CRU, while it is ~ 1 °C with respect to UoD. Therefore, any conclusion regarding simulated bias depends on the choice of reference data.

Using the above-mentioned model selection criteria, we find that 14 models exhibit both precipitation and temperature biases within the ± 1.5 STD range of observations. Given that these analyses are conducted using domain averages, it is quite possible that the low bias of some of these selected 14 models may be due to compensating errors across the analysis domain. Therefore, the selected models are further evaluated to determine how well their spatial distributions of temperature and precipitation agree with the corresponding observed patterns (Fig. 7).

Fig. 7
figure 7

Scatter plot between pattern correlation coefficient (PCC) and root–mean–square error (RMSE) for precipitation (a, b) and temperature (c, d) for two observation datasets and 14 selected CMIP6 models averaged over the entire domain for the period 1995–2014. The vertical dashed line shows the threshold of RMSE (1 mm/day for precipitation) and (2 °C for temperature). The horizontal dashed line shows the PCC thresholds 0.60 for precipitation and 0.96 for temperature. Models with open (white) circles exceed the error thresholds for selection, while the closed (black) circles indicate the candidate models. The final selection of models is shown by an asterisk in Table 1

Full size image

Figure 7 shows PCC and RMSE for annual mean precipitation and temperature for each of the selected models. These analyses are shown separately for each reference data (CRU, GPCC) to account for observational uncertainty. Additional imposed selection criteria include the RMSE ≤ 1 mm/day for precipitation (Fig. 7a, b), and ≤ 2 °C for temperature (Fig. 7c, d). The filled circles represent models that satisfy the new criteria, while the hollow circles represent those that do not. Of the 14 models, nine fulfill these conditions for both precipitation and temperature with respect to both sets of observations. These models have been marked with an asterisk next to their names in Table 1. The rationale for adopting this two-step model selection procedure is to retain only those models that demonstrate good fidelity in simulating the present climate, which should increase confidence in the reliability of future projections (Knutti et al. 2017). We also recognize that the selection of reference datasets has an impact on the statistics and that model validation based on the different reference datasets may result in a different selection of models. Such observation-based uncertainty is more pronounced in the case of precipitation, partly due to the fact that precipitation is quite variable in space and time, and its magnitudes are more sensitive to the choice of algorithms used in its interpolation from stations to grids. This is particularly true over topographically complex regions, such as NCA and SCA, as noted by Hannah et al. (2017). This disagreement among observational datasets has been discussed in previous works over these regions (e.g., Cerezo-Mota et al. 2016; Cavazos et al., 2020).

4 Future Changes in Annual Mean Temperature and Precipitation

Figure 8 shows the mean changes in annual temperature and precipitation for the near (2021–2040), mid (2041–2060) and far (2080–2099) futures with reference to the 1995–2014 period under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios. Given that radiative forcing is very similar across all future scenarios in the initial decades, the projected changes in near-future temperatures are not very sensitive to the choice of the emissions pathway (Fig. 8, left panels). These changes include an increase in temperatures over the northern U.S. of up to 2 °C and over Central America of ~ 1 °C. For the mid-future period (Fig. 8, middle column), the projected increase in temperature is up to 3 °C under SSP2-4.5 and 4 °C under SSP5-8.5 scenario. As expected, the strongest increase in temperatures is seen in the far-future over the entire domain under both SSP-2–4.5 and SSP5-8.5. These increases are particularly pronounced (+ 6 °C) over the northern half of the U.S. as a result of changes in the snow albedo feedback (e.g., Colorado-Ruiz et al. 2018; Ashfaq et al. 2016). Warming over the Caribbean is relatively low (+ 1.2 °C for SSP-2–4.5 and + 3.7 °C for SSP-5–8.5), as found by Vichot-Llano et al. (2020b). Table 2 summarizes the area-averaged changes in future simulations over the six subregions (see Fig. 1). The projected temperature changes are significant and robust as more than two-thirds of the models (66%) agree on the direction of the temperature change over the entire domain.

Fig. 8
figure 8

Spatial distribution of changes in annual mean temperature (°C) for the near (2021–2040), mid (2041–2060), and far (2080–2099) future, under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios with respect to 1995–2014 using the ensemble mean of M31. The backslash and forward slash indicate the grid boxes showing significant and robust changes, respectively, while hatching represents the grid boxes having both significant and robust changes. Significance is defined based on the two-tailed Student t test, while robustness is when at least 66% of the models project a climate change signal in the same direction

Full size image
Table 2 Projected changes in mean annual temperature (Temp, °C) and precipitation (Prec, %) over six subregions for near (2021–2040), mid (2041–2060) and far (2080–2099) future periods with reference to the present climate (1995–2014) under three SSP scenarios based on the M9 ensemble in Table 1 (GCMs with an asterisk in Table 1)
Full size table

As compared to the present climate, the projected annual precipitation displays a wetter response over the contiguous U.S. and drier response over Central America under the SSP2-4.5 and SSP5-8.5 forcing throughout the twenty-first century (Fig. 9), which progressively increase in magnitude over time as a function of the strength of the radiative forcing. By the end of the twenty-first century, precipitation is projected to increase (decrease) by 10–30% (10–40%) over the U.S. (Central America) under SSP5-8.5 scenario. The change in precipitation during the winter and spring is consistent with the annual pattern (Fig. S2). There is little variation in the SCA and CAR subregions in spring, although the precipitation response is relatively mixed in summer and autumn. For instance, by the end of the twenty-first century, change in precipitation is negative under the SSP5-8.5 scenario except over parts of WNA and ENA. Similarly, autumn precipitation is projected to increase with the exception of some pockets. The simulated precipitation response is less robust over CNA and NCA under the SSP1-2.6 and SSP2-4.5 scenarios than that under the SSP5-8.5 scenario. The north–south projected changes in precipitation are in agreement with the previous projections of a decrease in future precipitation over the southern subregions, including most Caribbean countries, using the SRES climate scenario A2 (Karmalkar et al. 2011), A2 and B2 (Campbell et al. 2010) and RCP4.5 and RCP8.5 (Maloney et al. 2014; Torres-Alavez et al. 2014; Colorado-Ruiz et al., 2018). However, the complexity of the Caribbean subregion requires further investigation, as previous studies suggest that there are marked differences in the precipitation trends in the present and future periods between northwestern and southeastern parts (Imbach et al. 2018; Taylor et al. 2018; Oglesby et al. 2016).

Fig. 9
figure 9

Same as Fig. 8 except for precipitation change (in %)

Full size image

Figure 10 shows the interannual evolution of temperature changes over each of the six subregions in the 2021–2099 period with reference to the 1995–2014 period in M31 under the three emission scenarios. After 2040, simulated changes under the three scenarios begin to diverge for most of the regions. As previously noted, the pace of warming is a function of the magnitude of the radiative forcing with the strongest increase under SSP5-8.5 and the slowest increase under SSP1-2.6. In fact, temperature shows a decreasing trend after the 2070s in all subregions under SSP1-2.6 as the radiative forcing gradually gets weaker after peaking before 2070 (Fig. 10). A detailed description of the uncertainty associated with the projected annual mean temperature over the six subregions under the three scenarios is provided in Table 3. The uncertainty in the trajectory of temperature changes is larger under the high-emission scenario than under the other two scenarios (Table 3). Moreover, regardless of the scenario, uncertainty increases towards the end of the century (shaded areas in Fig. 10). These results indicate that over the course of the twenty-first century, a large increase in temperature is likely over the U.S., while a relatively smaller but still significant increase in spread is projected over Central America and the Caribbean, particularly under the SSP2-4.5 and SSP5-8.5 scenarios. These findings are consistent with earlier studies (e.g., Colorado-Ruiz et al. 2018) and for the SCA and CAR regions (Hidalgo et al. 2013; Hidalgo et al. 2017; Vichot-Llano et al. 2020b) using CMIP5 models.

Fig. 10
figure 10

Projected evolution of changes in mean annual temperature in the twenty-first century with respect to 1995–2014 under SSP1-2.6, SSP2-4.5, and SSP5-8.5 using a M31 ensemble. Aqua, blue, and maroon curves represent the result for median values under SSP1-2.6, SSP2-4.5, and SSP5-8.5, respectively, and the shaded areas around each of the curves represent the likely range (66% of the projected changes). The curves are obtained by taking the difference of each future year with respect to the average from the historical period (1995–2014), and then smoothing with a 7-year running average

Full size image
Table 3 The uncertainty (66% likely range) associated with the projected changes in mean annual temperature (°C) and precipitation (%) over six subregions for the near (2021–2040), mid (2041–2060), and Far (2080–2099) future periods as compared to the reference period (1995–2014) under three SSP scenarios
Full size table

Figure 11 shows the projected evolution of the change in mean annual precipitation for the period 2021–2099 with reference to the period 1995–2014. As in Fig. 10, the curves represent the median of the 31 CMIP6 models, while the shading represents the likely range. As previously noted in the spatial maps, there is an increasing trend over the U.S. under all three scenarios. In contrast, the projected precipitation shows a decreasing trend over Central America and the Caribbean under SSP2-4.5 and SSP5-8.5, and little or almost no change under the low-emission scenario. As in the case of temperature, the spread of uncertainty among CMIP6 models (66% likely range) is large for the high-emission scenario as compared to the other two (Table 3). For all scenarios, the uncertainty spread increases towards the end of the twenty-first century (Fig. 11, Table 3). The temperature and precipitation trends for each of the six subregions over the period 2021–2099 under the three SSPs are shown in Table 4.

Fig. 11
figure 11

Same as Fig. 10 except showing precipitation changes (in %)

Full size image
Table 4 Trends of the projected annual temperature (°C per decade) and precipitation (% per decade) over six subregions for the period 2021–2099 under three SSP scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5. All trends are significant at the 95% significance level
Full size table

So far, the analyses of the projections of temperature and precipitation changes are based on the ensemble mean of the 31 CMIP6 models. To investigate the sensitivity of climate change signals over the study domain to the choice of models based on their accuracy over the reference period, we assess these projections using the mean of the 9 best-performing CMIP6 models (hereafter M9), identified in Table 1 with an asterisk, which have relatively small bias in the reference period. This reduced ensemble of selected models projects spatial changes in temperature and precipitation for the near, mid, and far futures under the three scenarios (Fig. S5) that are similar to those projected by the ensemble of all models (see Figs. 8 and 9). However, the magnitude of temperature changes is larger in the limited selection of models, particularly in the second half of the twenty-first century. Similarly, the selected models also produce relatively strong precipitation changes than the full ensemble in all time periods, and under all scenarios, with some minor exceptions. The seasonal changes in temperature and precipitation based on the M9 ensemble are shown in Figs. S6 and S7, which are very similar to the changes in the M31 ensemble (Figs. 8 and 9).

Table 2: Projected changes in mean annual temperature (Temp, °C) and precipitation (Prec, %) over six subregions for near (2021–2040), mid (2041–2060) and far (2080–2099) future periods with reference to the present climate (1995–2014) under three SSP scenarios based on the M9 ensemble Table 1 (GCMs with an asterisk in Table 1).

The time evolution of changes in individual months over the six regions using the M31 ensemble under SSP2-4.5 and SSP5-8.5 is shown in Fig. 12. Over the U.S., temperature increases are particularly pronounced in late summer and fall months after the 2050s under the SSP5-8.5 scenario. Over Central America, the increases under SSP5-8.5 are more pronounced in summer. As previously noted, the M9 ensemble projects temperature changes similar in direction to those of the M31 ensemble (Fig. 13), but the magnitude of increase is larger. Similarly, the changes in temperature under the low-emission scenario are relatively small in both the M31 and the M9 ensembles (Fig. S8).

Fig. 12
figure 12

Projected changes in monthly mean temperature (°C) over six subregions during the twenty-first century with respect to 1995–2014 under SSP2-4.5 (left panels) and SSP5-8.5 (right panels) using the M31 ensemble. The horizontal axis shows the years; the vertical axis shows the months

Full size image
Fig. 13
figure 13

Same as Fig. 12 except the changes are shown for the M9 ensemble

Full size image

For the period 2021–2099, temperature changes simulated by the M31 ensemble across all months over all six subregions are positive, except for 2 months under SSP1-2.6 and 1 month under SSP5-8.5. For the M9 ensemble, negative changes are projected for 12 months, 19 months, and 9 months under SSP1-2.6, SSP2-4.5, and SSP5-8.5, respectively, which account for 0.12%, 0.33%, and 0.16% of the total months under SSP1-2.6, SSP2-4.5, and SSP5-8.5, respectively.

Monthly changes in precipitation in the M31 ensemble exhibit characteristics similar to those for temperature when compared across the three future scenarios (Fig. 14). Increases in precipitation over the three U.S. regions occur predominantly in the winter months and intensify over time. The future precipitation deficit over NCA is mostly a result of drying in spring while that over SCA is due to drying in both spring and summer, which strengthens with increases in the radiative forcing. Moreover, in SCA and CAR, the driest months of the summer are July–August, which suggest an enhanced and extended MSD, especially after the 2050s and under the high-emission scenario. While the pattern of change is similar in the M9 ensemble (Fig. 15), the magnitude of change is stronger and more variable. Lack of substantial differences between the M31 and M9 ensembles suggests that the dipolar precipitation response in CMIP6 models is robust and does not necessarily depend on their accuracy in the reference period. Such an insensitivity of the simulated precipitation response to the performance-based selection of GCMs potentially suggests that the large-scale dynamic and thermodynamic responses to increases in radiative forcing in CMIP6 GCMs are perhaps dictated by similar synoptic-scale processes across the models that are relatively insensitive to regional biases over the terrestrial domain used in our analyses. However, further investigation is needed to quantify the underlying causes of the relatively robust precipitation response over these regions.

Fig. 14
figure 14

Changes in monthly precipitation (%) in each of the six subregions with respect to 1995–2014 under SSP2-4.5 (left panels) and SSP5-8.5 (right panels), using the M31 ensemble for the period 2021–2099

Full size image
Fig. 15
figure 15

Same as Fig. 14 except the changes are shown for the M9 ensemble

Full size image

5 Concluding summary

In this study, we analyzed 31 CMIP6 models to assess the projected temperature and precipitation changes over six subregions across the contiguous U.S., Central America, and the Caribbean during the twenty-first century under three shared socioeconomic scenarios (SSP1-2.6; SSP2-4.5, and SSP5-8.5).

We first examined the performance of the CMIP6 models over the reference period 1995–2014. The CMIP6 ensemble reproduces the observed distribution of temperature to within − 0.93 °C to 0.88 °C (− 0.02 °C to 1.27 °C) relative to CRU (UoD) over the full domain and reproduces the observed distribution of precipitation to within − 37.90–58.45% (− 33.23–51.96%) relative to CRU (GPCC) over much of the domain. However, models produce too much precipitation over the western U.S. and the monsoon region and too little over southern Central America.

We examined the projected changes in temperature and precipitation for three future periods (2021–2040; 2041–2060; 2080–2099) with reference to the 1995–2014 period under three future scenarios. The CMIP6 ensemble projects a continuous increase in temperature over the U.S., Central America, and the Caribbean under all three future scenarios. The temperature increase is more pronounced over the U.S. than over Central America and the Caribbean. By the end of the twenty-first century, the temperature is projected to increase by up to approximately 6 °C over northern parts of the domain under the high emission SSP5-8.5 scenario.

The spatial distribution of precipitation changes reveals a meridional dipole-like pattern, with an increase (10–30%) over the U.S. and a decrease (10–30%) over Central America and the Caribbean under all three scenarios with regional and seasonal variations over the entire domain. The eastern and central subregions (e.g., ENA and CNA) show an increase in precipitation during winter and spring, and a decrease during summer and autumn. The projected precipitation over Central America exhibits decreases during the winter, spring, and summer seasons, and an increase in autumn. The projected changes in precipitation are more pronounced towards the end of the twenty-first century. Even though there is no clear trend in precipitation changes for some regions, such as SCA, the positive temperature trends can increase aridity and decrease runoff through the intensification of regional hydrological cycle (Alfaro-Córdoba et al. 2020; Moreno et al. 2019; Veas-Ayala et al. 2018).

We further investigated projected changes using a subset of 9 of the best-performing models. The ensemble mean of selected models, does not exhibit significant differences in projected changes when compared with the ensemble mean of all models. Further investigation of the dynamic and thermodynamic characteristics is needed to quantify the relative insensitivity of simulated future responses to biases in the reference period. The use of two observational datasets for model evaluation highlights the uncertainty that exists in the observations and underscores the value of using multiple observational datasets in model evaluation.

The projected changes are spatially consistent between CMIP6 and CMIP5 models over this region. There are differences in terms of magnitude of the projected changes, which is understandable given the differences in the strength of radiative forcing between the SSP/RCP scenarios of CMIP6 and the RCP scenarios of CMIP5. Interestingly, autumn is the only season that exhibits positive rainfall anomalies in the southern tropical subregions in both the CMIP5 and CMIP6 (NCA, SCA), reflecting the possibility of delay in the demise of the summer precipitation season and the monsoon rains (e.g., Ashfaq et al. 2020; Bukovsky et al. 2015; Colorado-Ruiz et al., 2018; Lee and Wang et al. 2014; Torres-Alavez et al. 2014). Although we note that large-scale GCM-based projected changes in precipitation and temperature are less uncertain over the study region, fine-scale processes are known to influence the distribution of these changes at local scales (Diffenbaugh et al. 2005; Ashfaq et al. 2016), particularly over regions with complex topography. In addition, with respect to Caribbean small islands, it is important to note that some islands are not resolved at the spatial resolution of some of the GCMs (Campbell et al. 2010) used in this study. Therefore, further downscaling of CMIP6 GCMs is recommended before using them to quantify climate change impacts on natural and human systems across the U.S., Central America and the Caribbean.