We used daily maximum temperature at 2 m (T2m) from the ERA5 reanalysis over SSA ([25, 60]° S and [80, 40]° W) with a regular 2.5° resolution during the austral warm seasons (October–March) of 1979–2023 (Hersbach et al., 2020). We also employed data from 26 GCMs of the Climate Model Intercomparison Project Phase 6 (CMIP6, see Table S1 in the Supplement for details). To maximise the number of models, we considered one ensemble member per GCM. Although this strategy does not remove internal variability – an issue that may require large ensembles (e.g. Deser et al., 2020, and references therein) – it does increase the sample size available for constructing the storylines (requiring three or more ensemble members per model would have reduced the ensemble to less than half its original size). Daily maximum near-surface (2 m) air temperature (TX) was used for the definition of extreme temperature indices. In addition, monthly fields of sea surface temperature (SST), soil moisture content (SM, within the top 0–10 cm of the soil), 500 hPa geopotential height (Z500) and outgoing longwave radiation (OLR) were employed for the construction of the drivers (see Sect. 2.3). GCM historical simulations (Eyring et al., 2016) over the period 1979–2014 and Shared Socioeconomic Pathway projections (SSP5-8.5, O'Neill et al., 2016) for the 2015–2099 period were obtained from the CMIP6 archive. A common 2.5° × 2.5° horizontal grid and the austral summer season (December–January–February, DJF) were considered for both reanalysis and GCM simulations. Bilinear interpolation was used for TX, SST, Z500 and OLR data, while a conservative remapping was applied to SM data to avoid spurious values (Jones, 1999).

For most of the analyses, extreme temperature conditions are diagnosed based on the summer maximum of TX (TXx). This index emphasizes the magnitude of extreme events, rather than their frequency or duration, assuming that extremes occur every summer. TXx is computed at each grid point, and at regional scales, using the regions defined in the next section. Regional TXx was calculated by first averaging TX over the region and then selecting the maximum value for each summer in order to ensure warm widespread conditions at the regional level.

To identify spatially coherent regions, we followed the clustering procedure of Suli et al. (2023). Herein, the identification of homogeneous regions is based on clustering grid points with a high co-occurrence of local temperature extremes. To do so, we identified extremely warm days at each grid point as sequences of at least three consecutive days in which T2m exceeded the local daily 90th percentile of the 1981–2010 baseline period, using a 31 d moving window. Then, we applied the bottom-up Ward's hierarchical clustering method (Ward, 1963) to identify land grid points with a high co-occurrence of extremes (see Sect. 2.2 of Suli et al., 2023, for further details). As a result, five climatologically homogeneous regions were identified in SSA, which are consistent with those obtained from station-based data in Suli et al. (2023). The identified regions are depicted in Fig. 1, and named as northern SSA (NS), central-eastern SSA (CES), central Argentina and northern Argentinian Patagonia (CA), central Chile (CCH), and southern SSA (SS), including Argentinian Patagonia and southern Chile.

To ensure consistency in the spatial analysis, the same SA regions of Fig. 1 were also applied to each CMIP6 GCM. However, the CCH was excluded from the analysis due to substantial temperature biases associated with unresolved topography in GCMs, which can reach magnitudes of up to ∼ 8 °C in northern Chile (Salazar et al., 2024). Note that this regionalisation aims to provide a robust characterisation of regional extremes, rather than to identify areas of homogeneous changes or high uncertainty in future projections. The latter approach would maximise the ensemble spread at regional level but would also shift the focus away from the behaviour of spatially coherent regional phenomena and their underlying drivers.

For the regional analysis, the following drivers were considered using a hybrid approach that combines linear regression analysis (Sect. 3.2) with physical reasoning (see the Introduction section, and references therein). They are grouped into local and remote drivers. Local drivers are proximate factors that directly influence regional temperature, whereas remote drivers represent large-scale influences or teleconnections affecting the region:

Sea surface temperature in Niño 3.4 region (N3.4, remote driver): mean summer SST in the Niño 3.4 region ([5° N–5° S; 120–170° W]).

For each region, storylines describe the combined effect of the drivers' changes on summer TXx projections. Climate change responses, denoted as Δ, are computed for TXx and the drivers as the difference of the summer mean between the far future (2070–2099) and the historical period (1979–2014). The methodology used in this study follows the framework proposed by Zappa and Shepherd (2017) and is briefly described below. Firstly, we computed the climate change responses of the drivers for each region and each GCM. Secondly, the regional ΔTXx response was modelled separately for each region using an ordinary multi-linear regression (MLR, Eq. 1). In this regression, ΔTXx is the dependent variable (target), while two drivers act as independent variables (predictors): 1ΔTXxGW=ax+bx⋅ΔD1GWm′+cx⋅ΔD2GWm′. We only considered two drivers per region in order to limit the number of storylines (given by 2n, with n being the number of drivers). Limiting the selection to two drivers also facilitates the interpretation of the storylines and helps avoid overfitting in the MLR caused by interdependencies among the predictors. In Eq. (1), ΔD1 and ΔD2 represent the changes in the two drivers for each model m. The symbol (^′) indicates the standardised change relative to the multi-model mean (MMM), ax, bx and cx are the regression coefficients: ax denotes the MMM intercept, representing the expected mean response when there is no deviation in the driver responses relative to the MMM; bx and cx quantify the sensitivity of regional ΔTXx to each driver. Both the target (ΔTXx) and the drivers (ΔD1 and ΔD2) were scaled by global warming (GW) defined as the corresponding change in the area-weighted global mean near-surface temperature. The MLR is based on 26 values (GCMs) and was computed separately for each region.

Once the sensitivity coefficients were obtained, regional ΔTXx can be estimated for given values of GW and drivers' responses. Combining opposite (strong or weak) responses of the two drivers for each region results in four different storylines, which reflect the corresponding effects in ΔTXx. The final ΔTXx response follows Eq. (2): 2ΔTXxGW=ax+bx⋅t+cx⋅t. Here, t denotes the storyline index, which measures the magnitude of the driver responses (in standard deviations; SDs). In this case, the changes of the two drivers were selected to have equal standardised amplitudes, which also allows us comparing their relative effects in ΔTXx. As described by Zappa and Shepherd (2017), t was chosen to lie within the 80 % confidence region of the drivers' responses (see black stars in Fig. 4), which was obtained by fitting a bivariate normal distribution (t ∼ ±1.26 SD). Full details of the methodology can be found in Zappa and Shepherd (2017), in Appendix A of Mindlin et al. (2020) and in Garrido-Perez et al. (2024). In the construction of storylines, we assume that model biases remain constant in the future, and therefore do not substantially influence the climate change signals. Figure S3 supports this hypothesis by revealing no statistically significant relationship between model biases in TXx and their projected changes, ΔTXx. Furthermore, recent studies indicate that CMIP6 models reproduce the climatology and seasonal variability of the aforementioned drivers reasonably well over SSA when compared with ERA5, including challenging variables like SM (Qiao et al., 2022).

Figure 2 shows the MMM summer projections of TXx and the drivers used in this study for 2070–2099 (with respect to 1979–2014). Consistent with Almazroui et al. (2021a), tropical regions in northern SA exhibit a strong significant warming by the end of the century, exceeding 5 °C. In SSA, the largest TXx increases are projected along the Andes Mountains (Fig. 2a), aligning with Salazar et al. (2024), who reported a warming of up to 6 °C in northern Chile. In contrast, central SSA regions display a more homogeneous and less pronounced warming of approximately 4 °C (Lagos-Zúñiga et al., 2024).

Concerning the projected changes in the climate drivers, SM is expected to decrease significantly over northern SA, especially in the Amazon and along the Andes Mountains (see Fig. 2b), consistent with Cheng et al. (2017). In contrast, future SM projections for central-eastern SSA remain uncertain. In this region, CMIP5 GCMs projected positive SM changes for 2061–2080 compared to 2006–2025 under a high emission scenario (Cheng et al., 2017). However, more recent CMIP6 projections under SSP5-8.5 show no consistent changes in regional SM (Cook et al., 2020). As this region is characterised by strong soil-atmosphere coupling, the uncertainties in SM projections are expected to propagate to summer temperature changes.

Regarding SSTs, the central-eastern Pacific is projected to warm by up to 4 °C above historical values by the end of the century (Fig. 2c). This warming enhances convection over the region as it can be seen by negative changes in OLR (Fig. 2d). Pronounced warming is also observed in the western Pacific Ocean near southeastern Australia (Fig. 2c), as noted by other authors (Lenton et al., 2015; Oliver et al., 2014). Although the direct impact of the western Pacific Ocean warming on SA remains uncertain, Sun et al. (2023) suggested that air-sea coupling in the tropical Pacific greatly amplifies the atmospheric response of the South Pacific to ENSO. Indeed, ΔZ500^* exhibits alternating anomalies over the Pacific that resemble a Rossby wave pattern extending from southern Australia (Fig. 2e), which has been associated with heatwaves in the subtropical SA (Cerne and Vera, 2011; Shimizu and de Cavalcanti, 2011). In addition, enhanced anticyclonic conditions are projected in southern SA, particularly at high latitudes, which may be linked to an increasing zonal asymmetry of the Southern Annular Mode during DJF (Campitelli et al., 2022).

To identify remote and local drivers of ΔTXx, we examined the regression patterns of several variables and constructed indices displaying a strong and physically consistent relationship with regional TXx. Figure 3 illustrates the linear regression patterns of these field responses onto regional ΔTXx. Figure 3a shows significant positive SST regression coefficients over the tropical Pacific (yellow box) suggesting that enhanced El Niño events contribute to increase ΔTXx in NS. Although El Niño is currently associated with cooler TXx conditions in this region compared to La Niña (Arblaster and Alexander, 2012), future projections suggest a weakening of the ENSO-related temperature signal over SSA (McGregor et al., 2022). Consequently, El Niño events may exert a reduced cooling effect in the future, resulting in higher TXx values relative to the present, and thus contributing to a positive ΔTXx response. In addition, changes in the OLR gradient (ΔOLRg) act as an important remote driver of NS ΔTXx (Fig. 3b). In particular, enhanced mid-latitude convection (negative ΔOLR in the poleward box) and/or suppressed subtropical convection (positive ΔOLR in the equatorward box) is associated with amplified warming over this region.

For the other regions, ΔOLRg also acts as a remote driver of ΔTXx in CES, where regional warming concurs with an anomalous configuration of the subtropical Atlantic anticyclone, or with modified SACZ-related convection (Suli et al., 2023). Furthermore, projected drying over northern Argentina and Paraguay (green box in Fig. 3c) is consistent with a CES warming response, although significance is limited to few points, possibly reflecting a weakened soil–atmosphere coupling under future climate conditions (Ruscica et al., 2016). Regarding CA (Fig. 3e–f), the results show that GCM projections with larger decreases in SM_CA or an enhanced OLRg display more pronounced TXx warming. Finally, the largest warming in SS (Fig. 3g–h) is associated with reduced SM_SS and an anomalously high Z500^* over SSA. The influence of SM_SS is consistent by Collazo et al. (2024), who found that southern SA exhibits strong soil-atmosphere coupling during the warm season, despite its aridity. Likewise, the Z500HL* driver captures high-latitude blocking, which has been linked to SS heat extremes (Suli et al., 2023). Previous studies also indicate that anticyclonic anomalies over southern SA can trigger heatwaves in this region (Cerne and Vera, 2011; Collazo et al., 2024; Jacques-Coper et al., 2016).

In the following, different MLR models (see Eq. 1), based on the climate change responses of TXx and the regional drivers described in Sect. 2.3 were performed for each SSA region. Sensitivity tests were also conducted to assess whether lagged relationships between the changes in the drivers and regional TXx could improve the model performance. However, no significant improvement was found when introducing temporal lags. Therefore, we focused on simultaneous summer responses in both drivers and TXx. We also verified that the selected drivers were not significantly correlated with each other (i.e. Pearson correlation coefficients with p values > 0.1) to avoid redundant information that would add unnecessary complexity to the model and the interpretation of the drivers.

The final combination of drivers is outlined in Table 1, along with the sign of their regression coefficients (+/-) and the corresponding explained variance (R2). The selected drivers vary across SSA regions. In NS, the warming response in TXx is substantially affected by changes in remote drivers (ΔN3.4 and ΔOLRg), while in SS only local drivers are identified (ΔSM_SS and ΔZ500HL*). In contrast, both local and remote drivers (ΔSM_i and ΔOLRg) affect the warming of extremes in CES and CA. For all regions, the uncertainty in the drivers' changes is significantly correlated with that in TXx, except for CES, where ΔSM_north does not show a significant response in ΔTXx. Although this region exhibits a strong soil-atmosphere coupling on interannual timescales (Jung et al., 2010; Ruscica et al., 2015; Sörensson and Menéndez, 2011), the lack of significance indicates that SM_north cannot explain the spread of TXx projections in this area. For all regions, R2 exceeds 35 %, with the highest values observed in SS (R2 ∼ 41 %) and CA (R2 ∼ 56 %).

Four storylines (herein labelled as ST#, with # ranging from 1 to 4) of future summer TXx changes were constructed based on the combination of the two most influential drivers of ΔTXx in each region described in Sect. 3.2. Figure 4 depicts the scatterplots of the two drivers' responses within the CMIP6 ensemble, along with the standardised change amplitudes selected to construct each storyline (represented by black stars), following the regression framework described in Sect. 2.4. GCMs that displayed a systematic outlier behaviour across all regions were excluded from the analysis. The results show considerable uncertainty in driver responses. Some drivers show consistent changes in sign but have uncertain magnitudes (e.g., N3.4 and the SM_SS index, see Fig. 4a and d), while for others, both the sign and magnitude of the change are uncertain (e.g., SM_CA or OLRg, see Fig. 4c). For all drivers, the spread of responses across the multi-model ensemble is clearly distinguishable from the internal variability. To illustrate this, we compared the magnitude of the projected changes in the selected drivers with an estimate of the internal variability based on the interannual standard deviation of the detrended series, following the approach in Mindlin et al. (2020, their Appendix B). The results (Table S2) show that the variability within individual models is significantly lower than across models (inter-model variability), indicating that the spread in driver projections is primarily driven by model uncertainty rather than by internal variability.

Each storyline characterises the summer ΔTXx as the combined response in bx and cx (see Eq. 2), yielding distinct patterns of warming depending on how the two drivers change. For instance, ST1 for the NS region (Fig. 4a and first row of Table 2) is characterised by lower-than-MMM changes in both N3.4 and OLRg, while ST4 represents the opposite pattern. Similarly, ST2 and ST3 correspond to opposing changes in these two drivers. The specific combination of drivers for each ST and region, as obtained from Fig. 4, is summarised in Table 2.

Figure 5 illustrates the scaled summer ΔTXx for each GCM (coloured circles), including the MMM (black diamond) and its 1 SD range (grey shading), as well as the reconstructed storylines of ΔTXx (coloured squares) based on the combination of drivers' responses. This figure also reveals which combination of drivers leads to the largest and smallest warming in each region, i.e. the worst-case and best-case scenarios, respectively. Overall, TXx warming appears unavoidable, as even the best-case scenario shows an increase of ∼ 0.9 K K⁻¹ across all regions. Moreover, the TXx warming response in the worst-case storyline is 29 % to 54 % higher than in the best-case scenario. To measure the robustness of the storylines, we compared the difference between opposite storylines (Fig. 5) with the median absolute error (MdAE) of the MLR (Table 1) for each SSA region, similar to Mindlin et al. (2020). In most regions (NS, CA and SS), the MdAE represents less than ∼ 25 % of the storyline responses. In CES, the MdAE represents a higher fraction of the differences between storylines (∼ 35 %), which may be due to the lack of significance of one of the drivers. Overall, these results confirm that the regression-based framework provides a meaningful representation of the TXx responses across SSA. Indeed, the inter-storyline variability reasonably encompasses the range of uncertainties in ΔTXx projections, represented by the grey-shaded areas in Fig. 5a–d. In the remainder, the two remaining storylines will not be discussed, as they exhibit an intermediate result.

In NS, the largest warming in TXx (ST4, purple square; Fig. 5a) results from the combination of a warming in the tropical Pacific and a strengthening of the OLRg relative to the MMM changes. This storyline determines a 54 % greater increase in ΔTXx compared to the opposite combination of drivers' responses. Differenly, Mindlin et al. (2024) found a larger increase in October–April mean TX over southwestern SA under a low Pacific warming storyline (i.e. a relative cooling of both eastern and central El Niño), whereas a high Pacific warming produced the opposite response. The discrepancies with our findings may stem from differences in the selected drivers, target variables, seasonal definitions, and/or regional domains. CES and CA storylines are constructed using similar drivers (as seen in Table 1). However, the response in ΔTXx differs between the two regions. For CES, ΔOLRg is the only driver with significant influence on ΔTXx (Table 1; Fig. 5b). This is reflected in the separation between the storylines. ST1 and ST2 are associated with a weakening of the OLRg, whereas ST3 and ST4 correspond to a weak intensification of the OLRg (Fig. 4b and Table 2), with the latter yielding an additional ∼ 29 % increase in ΔTXx. Comparatively, the difference in ΔTXx between ST1 and ST2, as well as between ST3 and ST4, is negligible. This pattern highlights that the spread of ΔTXx projections over CES is primarily driven by OLRg variations, likely linked to changes in the subtropical Atlantic anticyclone or in SACZ-related precipitation, while ΔSM_north does not play a significant role. In contrast, in CA, the combination of strong drying and an intensification of the OLRg leads to a ΔTXx warming ∼ 44 % higher than that associated with the opposite storyline (Fig. 5c). Finally, the storyline characterised by the largest warming in SS ΔTXx (ST3, blue square; Fig. 5d) is determined by the combination of enhanced drying and anticyclonic activity relative to the MMM, with an additional 30 % increase in ΔTXx warming compared to the best-case storyline (ST2, green square).

For better interpretation of the differences in the storylines of ΔTXx, Fig. 6 shows the composite difference of the spatial patterns of ΔTXx (shading) and ΔZ500^* (contours) between the GCMs following the worst- and best-case storyline of regional ΔTXx (i.e. GCMs falling within the 80 % confidence interval of the corresponding quadrant in Fig. 4). Enhanced warming under the worst-case outcome is evident across all regions, particularly in NS. In contrast, in CES, the ΔTXx differences between extremal storylines are small, consistent with the poorer MLR performance and the weak influence of one of its drivers. The worst-case scenarios of each region are also accompanied by distinctive circulation anomalies, featuring Rossby wave trains with different pathways and latitudes depending on the region, which are consistent with the enhanced regional ΔTXx responses. Likewise, in NS, CES, and to a lesser extent CA, the drivers associated with the largest warming in ΔTXx lead to anticyclonic anomalies over the South Atlantic Ocean. This is consistent with Suli et al. (2023), who found that heatwaves in these regions are triggered by shifts/intensification of the subtropical semi-permanent high-pressure systems. The influence of the subtropical anticyclone is missing in the worst-case storyline of SS, where extremely warm days are related to co-located anticyclonic anomalies (blocking) and jet meandering.

Overall, these findings underscore the importance of identifying region-specific drivers and exploring physically plausible scenarios beyond the MMM. For all regions, we find that changes in both thermodynamic (regional SM, N3.4) and dynamic (Z500HL*, OLRg) drivers contribute to the spread of future projections in regional ΔTXx, stressing the importance of understanding dynamical aspects of climate change in the region.