Introduction

Extreme heat events (EHEs) have profound socioeconomic and environmental impacts, including threats to human health, increased wildfire frequency and intensity, reduced food production, decreased productivity, and damage to public facilities1,2,3. Eastern China, with its dense population, is particularly vulnerable to these events4, making it a critical region in which to study their mechanisms and consequences. For example, in the summer of 2022, eastern China experienced the most intense EHE since 1951, with over 300 million people experiencing temperatures exceeding 40 °C and resulting in economic losses of 5 billion dollars5,6,7. Therefore, improving our understanding of the variability and drivers of EHEs is essential for developing effective mitigation and adaptation strategies.

The formation of EHEs is closely linked to the development of the so-called “heat dome” (HD), an atmospheric anticyclone that traps heat near the surface3,8,9. Persistent large-scale HD is capable of causing surface warming through adiabatic subsidence and incoming solar radiation10,11,12, creating favorable conditions for extreme temperatures. In East Asia, HD is strongly associated with changes in the Western Pacific Subtropical High (WPSH) and the South Asian High (SAH)6,13,14. The eastward extension of the SAH promotes sinking air motions, which in turn push the WPSH further west and enhance the likelihood of EHE conditions in eastern China15,16. Wang et al. 12 further highlighted that the key region of EHE is located in the southern part of HD, where easterly winds transport low-enthalpy air into eastern China, leading to anomalous subsidence and extreme heat. Numerous remote factors can contribute to the HD and EHEs in eastern China, especially the mid-latitude teleconnections17 due to internal variability7,12,18, mid-latitude SST8,19,20, and signals from the Arctic Oscillation (AO), sea ice and Arctic climate change10,21,22. Recent studies7,23 have highlighted the role of Pakistan convective precipitation (PCP) as key heat source that triggers downstream EHEs through a Rossby wave train propagating toward eastern China, which appears to be largely independent of ENSO23.

Academic and societal interest in EHEs has grown in recent years, driven by their increasing impacts2,3. Since the 1950s, heatwaves have become more frequent, longer-lasting, and more intense worldwide2,17,24,25. In particular, China has experienced a marked increase in the occurrence and severity of summer EHEs in recent decades11,26,27,28. Besides the extreme summer of 2022, notable EHEs also occurred in 2013, 2017, and 20186. While global warming is a key driver of this trend6, studies suggest that natural variability remains a key contributor, as seen in 202212,29,30. Worse still, projections indicate that EHEs will become more frequent in a warming climate, posing escalating challenges for China6,29,31. Therefore, understanding the change and mechanism of EHEs becomes an urgent scientific priority.

EHEs arise from temperature anomalies that deviate from the mean state, with the magnitude of these anomalies reflecting changes in temperature variability. Therefore, temperature variability plays a crucial role in the occurrence of EHEs32. The frequent occurrence of EHEs in eastern China in recent years implies a potential shift in temperature variability. However, the mechanism of this change and its manifestation in the future warming climate remain poorly understood. Global warming is already altering the variability of some climate variables, with profound implications for human society33. Nevertheless, the influence of shifting variability in eastern China summer temperature anomalies on the frequent occurrence of local EHEs has received less attention. Consequently, projections of global warming on the concurrent climate change in local EHEs are inherently uncertain.

Here, we investigate the recent changes in high-summer (July–August) interannual temperature variability in eastern China, with particular attention to the south-north uniform-warming (SNUW) mode and the south-north meridional dipole (SNMD) mode. We aim to clarify whether large-scale temperature modes have changed, and to identify what physical mechanisms may drive these changes. Projections from climate models are further analyzed to distinguish the roles of natural variability and anthropogenic warming and to explore the potential future evolution of these interannual temperature variability modes under global warming.

Results

Reversal in variability of leading temperature modes

High summer temperatures in eastern China tend to be higher and more variable than those in neighboring areas (Supplementary Fig. 1a, b). Here we apply the empirical orthogonal function (EOF) analysis to two-meter temperature (T2m) anomalies during 1984–2023 (see Methods), which identifies two dominant modes that explain significantly higher variance than the remaining modes (Supplementary Fig. 1c). Therefore, this study focuses on these two modes. The first is the SNUW mode, characterized by positive temperature anomalies centered along the Yangtze River (Supplementary Fig. 1d). The second is the SNMD mode, featuring opposite temperature anomalies between northern and southern regions of the Yangtze River (Supplementary Fig. 1e). Together, these two modes explain ~70% of the total T2m variability (Supplementary Fig. 1f–h). The SNUW and SNMD indices, defined by regional-mean T2m (Fig. 1a, see Methods for definitions), exhibit strong Pearson correlation coefficients reaching 0.97 and 0.92 with the corresponding time series of EOF modes (Supplementary Fig. 1i), respectively. Thus, the SNUW and SNMD modes effectively capture the key features of the summer temperature in eastern China.

Fig. 1: Reversal of summer temperature modes in eastern China.
Fig. 1: Reversal of summer temperature modes in eastern China.
Full size image

a Time series of standardized SNUW index (red), SNMD index (blue) during 1984–2023. Composite of anomalous T2m in SNUW years (b) and SNMD years (c), respectively. Years are listed in Supplementary Table 1. Boxes are used to define SNUW (box in c) and SNMD (box in d) indices. “SR” denotes the spatial correlation coefficient with the corresponding EOF modes. d 11-year sliding standard deviation of SNUW (red) and SNMD (blue), with the dashed line showing linear trends. Regression pattern (doubled in values) of empirical orthogonal function (EOF) for the second mode (EOF2; e) and first mode (EOF1; f) during 1984–2003, for the EOF1 (g) and EOF2 (h) during 2004–2023. Hatching suggests that anomalies pass the 95% confidence level of Student’s t test. Inside the parenthesis is the explained variance. Rectangles in (e–h) are the EOF region (22°–38°N, 105°–123°E).

Furthermore, in order to assess the ability of SNUW and SNMD indices to capture spatial patterns and the physical reality of EOF modes, we further conduct composite analysis for extreme SNUW and SNMD events that exceed 1.0 standard deviations for composite analysis (see Supplementary Table 1). In Fig. 1b, c, the composite results show highly consistent spatial characteristics with two EOF modes, respectively, with both spatial correlation coefficients reaching above 0.97 (p < 0.01). We also examined the T2m anomalies during the most extreme positive and negative anomaly years (Supplementary Fig. 2a–d), which also well reflect the EOF mode features. Additionally, excluding some extreme years leads to consistent EOF modes (Supplementary Fig. 2e, f). These results suggest that the SNUW and SNMD modes are physically meaningful in multiple years rather than mathematical artifacts.

The increasing dominance of the SNUW mode has become evident in recent decades. Accompanied by stronger fluctuations in SNUW anomalies after 2010, more EHEs (e.g., 2013, 2017, 2018, and 2022) have occurred in eastern China with a peak in 2022 (Fig. 1a, Supplementary Fig. 1i). In contrast, the SNMD index peaked in 2003, after which its variability markedly weakened, indicating a shift in the relative importance between the SNUW and SNMD modes. We further confirm this shift by a 11-year sliding standard deviation, which reveals a reversal in the variability of the two modes around 2004 (Fig. 1d), remaining consistent even with the long-term trend removed (Supplementary Fig. 3). To statistically identify the shift timing, the difference between standard deviation of two modes is calculated (Supplementary Fig. 4). A 11-year sliding Student’s t test, which represents the changes in interannual variability, suggests that the shift occurred in 2004. The emergence of this shift appears to result from two concurrent processes: an increasing SNUW variability after the early 2000s and an abrupt weakening of SNMD variability around 2004. Further explanation of their variability changes will be presented later.

To examine this change in details, we conducted additional EOF analyses for two sub-periods: 1984–2003 (P1) and 2004–2023 (P2) (Fig. 1e–h). During P1, the SNMD mode was dominant, explaining 39.3% of the total variance, while the SNUW mode accounted for 29.4%. In contrast, in P2, SNUW became the dominant mode, with its explained variance increasing to 44.7% and associated with stronger temperature anomalies (Fig. 1e vs. 1g). Meanwhile, the SNMD mode weakened significantly, with its explained variance dropping to 18.9%, and its characteristic cold anomaly over the northern China largely disappearing (Fig. 1f vs. 1h). Therefore, our results reveal a considerable reversal around 2004 in the interannual variability of dominant modes of summer temperature in eastern China, which played a crucial role in the increasing frequency of EHEs in this region.

Increase in SNUW variability

Changes in climate modes in eastern China are closely related to large-scale circulations and external forcings34. Therefore, we next focus on the potential changes in key factors affecting the summer temperature in eastern China. As shown in Fig. 2a, the changes of temperature in eastern China are closely related to the presence of an anomalous HD anticyclone over the northern China, which is favorable for the northeastward extension of the SAH and the EHE environments. A strong correlation is found between SNUW and HD indices (r = 0.87, p < 0.01; Fig. 2b). On a broader spatial scale (Supplementary Fig. 5), the anomalous atmospheric circulation associated with SNUW closely resembles the pattern excited by the PCP (see ref. 7). PCP can significantly trigger anomalous anticyclone over eastern China7,23, establishing strong correlations with both HD and SNUW indices (Supplementary Figs. 6a, b, 7). Additionally, the AO, which reflects oscillations in air pressure between the Arctic and the mid-latitudes35 (Supplementary Fig. 5), also plays a key role in modulating SNUW variability through the downstream-propagating wavetrain originating from the Arctic (see Methods, Supplementary Fig. 8). Notably, these modulations have intensified in recent years (Supplementary Fig. 6c, d).

Fig. 2: Intensified variability of SNUW mode associated with PCP and AO.
Fig. 2: Intensified variability of SNUW mode associated with PCP and AO.
Full size image

a Regression pattern of 200-hPa geopotential height (shading) and horizontal wind (vector) onto SNUW index. Red contours are the geopotential height climatology of South Asian High. Hatching suggests passing the 95% confidence level. Box outlines the region for heat dome (HD) index. b Time series of SNUW (red) and HD (blue) index. c 11-year sliding correlation coefficient between HD and SNUW (c1, green), between HD and PCP + AO (c2, red), between SNUW and PCP + AO (c3, blue). Bolded lines indicate passing the 90% confidence level. d 11-year sliding standard deviation of SNUW (red), HD (blue), PCP + AO (purple), PCP (green) and AO (black).

The combination of PCP and AO (denoted as PCP + AO; weighted by their regression coefficients, see Methods for details) is used to explore their joint association with SNUW. Since the correlation between PCP and AO is insignificant (R = 0.04), their effects can be regarded as simple linear addition. The mid-to-upper tropospheric anticyclones associated with PCP + AO exhibit a barotropic structure (Supplementary Figs. 6a, c, 9a) and are accompanied by the northeastward shift of SAH and the northwestward shift of WPSH. These shifts promote subsidence and reduced precipitation (Supplementary Fig. 9b, c). The co-occurrence of drought and high temperatures can further intensify EHEs22. PCP + AO explains over 30% of the interannual variations in temperature anomaly (Supplementary Fig. 9d), and this linkage has become increasingly pronounced in recent years (Fig. 2c). From 1984 to 2023, both PCP variability and AO variability exhibit increasing trends (Fig. 2d), coinciding with the increasing correlation coefficients between PCP + AO and SNUW, and between PCP + AO and HD. Notably, the significant (p < 0.05) upward trend of PCP + AO variability aligns well with the trends in SNUW and HD, yielding robust correlation coefficients of 0.83 and 0.89, respectively. These relationships explain ~68.9% and ~79.2% changes in the SNUW and HD variability changes, as evaluated by R2 (Fig. 2d). These findings underscore that the increasing variability of PCP and AO has potentially driven the recent rises in HD and SNUW variability. We also calculated the regional-mean (same as EOF region) explained variances of T2m sliding standard deviation by PCP + AO, which is 30.1%. Although both AO and PCP have increase in their variability, their contributions are not identical, with AO having a greater role in the 1984–2023 period. This can be seen both in their weights (i.e., 0.4 for PCP and 0.6 for AO, see Methods) and in their trends in variability, where the AO variability is more similar to that of SNUW (Supplementary Fig. 10). The increase in variability for AO is more dramatic and closely aligns with SUNW variability changes; whereas the increase in variability for PCP is relatively modest.

Decrease in SNMD variability

The decline in SNMD variability shows a clear abrupt shift around 2004 (Fig. 1d), potentially coinciding with a significant decrease in the summer sea-ice concentration (SIC) over the Barents-Kara region (Fig. 3a, b). To confirm the shift point, we define a Barents-Kara SIC index (see Methods and Fig. 3a). The SIC index shows an abrupt shift in 2004 based on a sliding Student’s t test (Fig. 3b). The local SIC declined by approximately 60%, from an average of 0.294 during P1 to 0.117 during P2. The sea-ice loss also led to a substantial reduction in its variability, with the same shifting point around 2004 (Fig. 3c). The strong correlation between SIC and SNMD variability (R = 0.77, p < 0.05) suggests that Barents–Kara sea-ice loss and its reduced variability accounted for ~60% (assessed by R2) of the SNMD variability reduction. While ~26.9% of the local T2m sliding standard deviation variance can be explained by SIC. Additionally, this shift also directly led to the reversal of SNUW and SNMD variability around 2004.

Fig. 3: Weakened SNMD mode impacted by Barents–Kara sea-ice loss.
Fig. 3: Weakened SNMD mode impacted by Barents–Kara sea-ice loss.
Full size image

a Interdecadal change in sea-ice concentration (SIC) between P2 (2004–2023) and P1 (1984–2003). Sector outlines the region for SIC index. b SIC index (green) with period average (horizontal line, P1 in red and P2 in blue) and 11-year sliding t-test value (black). c 11-year sliding standard deviation of SIC index (red) and SNMD index (blue). d regression pattern of SIC onto SNMD index during P1 (left) and P2 (right). R marks the corresponding correlation coefficient between SIC index and SNMD index. Regression pattern of 200-hPa eddy geopotential height (shading) and wave activity flux (vector) onto SIC index during P1 (e) and P2 (f). AO signal is removed from SIC index. Dots suggest passing the 0.05 significance level.

A comparison between P1 and P2 reveals a significant weakening of the SNMD-SIC relationship after 2004 (Fig. 3d). During P1, the correlation coefficient between SNMD index and SIC index reached 0.72 (p < 0.01) but it dropped to only 0.24 (p > 0.1) in P2. This decline was most pronounced in the northern sector (outlined in black in Fig. 3d), which corresponds to the region experiencing the most severe sea-ice loss (Fig. 3a). During P1, the SIC was linked to anomalous northern low pressure and southern high pressure circulations over eastern China (Fig. 3e; outlined by red rectangle) via a Rossby wave train, producing dipole-type circulation and temperature anomalies (Supplementary Fig. 11a–c). This wave train mechanism is consistent with previous works that Barents-Kara SIC drives downward Rossby wave train to modulate eastern China precipitation dipole in CMIP6 models36,37,38.

During this period, the SIC explained 27.21% of the temperature variability (Supplementary Fig. 11d). However, as sea-ice declined, the significant downward-propagating wave train disappeared (Fig. 3f), and climatic anomalies linked to SIC on eastern China became negligible in P2 (Fig. 3f, Supplementary Fig. 11e–h). Note that here we have removed the AO signal from the SIC via linear regression to avoid confusion with physical processes associated with AO in the SNUW mode. We also checked the relationship between SIC and SNUW, where the correlations are insignificant during both periods (R = −0.30 during P1; R = −0.27 during P2; both p > 0.1).

To further quantify the linkage of PCP + AO and SIC with temperature variability, we applied linear regression to remove their signals and analyzed changes in the standard deviation of temperature modes (Fig. 4). For the SNUW mode, removing the PCP + AO signal (see Methods for details) substantially reduced the magnitude of variability amplification from +42.03% to +5.33% (Fig. 4a). Similarly, for the SNMD mode, removing the SIC signal lowered the magnitude of variability reduction from −36.24% to −10.56% (Fig. 4b). Notably, SNMD exhibited a negative correlation between its northern and southern regions during P1 (R = −0.40, p < 0.1), but this relationship disappeared and even reversed to positive in P2 (R = 0.41, p < 0.1, Fig. 4c), indicating the near-complete disappearance of the SNMD mode. These results further confirm the critical linkage of PCP + AO to the enhancing SNUW variability and the role of SIC decline in weakening SNMD variability.

Fig. 4: Contributions of modulators to change in variability of SNUW and SNMD modes.
Fig. 4: Contributions of modulators to change in variability of SNUW and SNMD modes.
Full size image

a SNUW standard deviation during 1984–2003 and 2004–2023, attached with relative change after signal removal (last column), and relative change between two sub-periods (last row). b Same as (a) but for SNMD standard deviation. c Correlation coefficient between north (SNMD-N) and south (SNMD-S) poles of SNMD during 1984–2003 (blue) and 2004–2023 (green), with SIC signal and remove SIC signal, respectively. Dashed lines are the 90% confidence level threshold. In (b, c), SIC signal is removed based on corresponding periods because the correlation between SIC and SNMD is insignificant after 2004.

Projections of SNUW and SNMD in CMIP6

To distinguish the contributions of natural variability and anthropogenic forcing, and to assess future changes in SNUW and SNMD variability, we employ the Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model ensemble (MME, see Methods). We first examine the spatial patterns of SNUW and SNMD in CMIP6 models. Results show that the CMIP6 models can generally reproduce the broad region-wide warming of the SNUW mode and the meridional dipole characteristics of the SNMD mode, with each model exhibits high spatial correlation with the observation (Supplementary Fig. 12). In terms of intensity, the SNUW mode exhibits relatively small errors near the center of the defined region (box), with only marginal variations observed at the periphery (Supplementary Fig. 12c). However, the SNMD mode demonstrates significant errors in both the northern and southern regions, showing weaker intensity than that in the observed data. This may indicate that the model fails to accurately capture the dipole structure. Further analysis reveals that most models perform poorly in simulating the dipole relationship, unable to reproduce the negative correlation. Therefore, the simulation results of the SNMD mode should be interpreted with cautions.

The MME mean is used to isolate externally forced signals by averaging out natural variability29. While the AO index in CMIP6-MME shows an upward trend (Supplementary Fig. 13a), its variability remains within the range of natural oscillations in both historical and future projections (Fig. 5a). This demonstrates that the observed increase in AO variability between 1984 and 2023 is primarily driven by the natural variability, which is also corroborated by observations (Supplementary Fig. 14a, c).

Fig. 5: Future possible intensification of variability in SNUW mode.
Fig. 5: Future possible intensification of variability in SNUW mode.
Full size image

19-year sliding standard deviation of AO (a), PCP (b), HD and SNUW indices (c) in CMIP6-MME. Historical simulation and SSP585 simulation are marked with different colors or backgrounds. d Relationship between the projected 19-year sliding PCP variability and SNUW variability among different models averaged in 2043–2082. Shading is the 95% confidence interval.

In contrast, while the PCP index remained within the range of natural oscillations before 1980s, it has shown a significant increasing trend since then (Supplementary Fig. 14b, d). This trend is also evident in the PCP variability itself, which has exhibited a robust rise in recent decades (Fig. 5b), consistent with observations (Supplementary Fig. 14b). Previous study have attributed this long-term increase in PCP variability predominantly to the anthropogenic warming33. Correspondingly, both HD and SNUW variability exhibits trend similar to PCP variability (Fig. 5c), with oscillation of small magnitude prior to 1980 followed by a marked increasing trend thereafter. These results validate the reliability of CMIP6 in simulating SNUW and related variables, enabling us to explore future changes in SNUW with confidence.

Since AO variability is projected to retain its periodic natural oscillations through the mid-to-late 21st century, the increasing variability of PCP is expected to emerge as the dominant driver of future SNUW variability (Supplementary Fig. 13c). Here we apply the inter-model relationship to understand future SNUW and PCP relationship since it can reveal the linear relationship between two variables in multi-models39. Inter-model analysis indicates that models with higher future PCP variability tend to predict greater SNUW variability (R = 0.58; Fig. 5d). However, when considering the joint effect of PCP + AO, the influence of AO may not always contribute positively to the SNUW variability change due to its intrinsic periodic oscillations (Fig. 5a). The inter-model relationship of PCP + AO against SNUW is weaker (R = 0.43; Supplementary Fig. 15). This also implies that future changes in SNUW variability cannot simply be assessed using the current PCP + AO calculations. Considering that the variability of the PCP is increasing in a warming climate (Fig. 5b), and that the role of the AO may become significant only when its variability is unusually large (but still small compared to the PCP; Fig. 5a, b versus Supplementary Fig. 13c), the effect of the PCP variability is likely to dominate future changes in SNUW variability. In addition, because the linkage among PCP, AO, and SNUW variabilities is likely to evolve over time, future studies may need to refine their methodologies in light of real-time observations.

The disappearance of SNMD mode is evident in the reversal correlation between the northern and southern SNMD regions (Fig. 4c). We find that the 11-year sliding dipole relationship of SNMD mode is also closely tied to the declining SIC variability (R = −0.84, p < 0.1, Fig. 6a), suggesting that sea-ice loss has played a key role in this shift. Future projections from the CMIP6-MME indicate a significant trend toward the negative phase of SNMD index (Fig. 6b). However, this trend is primarily driven by the faster warming in the northern part than in the southern part (Fig. 6c). The accelerated northern warming promotes greater spatial uniformity in regional temperature changes across eastern China, as evidenced by the significant increase in the future correlation coefficients in Fig. 6d. We also note that the variability of SNMD seems to increase after 2004 (Fig. 1d). However, because SNMD is likely to have disappeared since then (insignificant anomalies in the northern region in Fig. 1h), the sliding standard deviation trend in SNMD (0.34 decade−1) likely responds to the trend in SNUW (0.32 decade−1), which is the reflection of increasing local temperature variability.

Fig. 6: Disappearance of SNMD mode.
Fig. 6: Disappearance of SNMD mode.
Full size image

a 11-year sliding correlation coefficient between north and south poles of SNMD index in ERA5 (blue; regions outlined in Fig. 1c) and sliding standard deviation of SIC index (red). b SNMD index in CMIP6 multi-member ensemble. c Time series of north (red) and south (blue) poles of SNMD index. d 19-year sliding correlation coefficient between north and south poles of SNMD index in CMIP6 multi-member ensemble. Historical simulation and SSP585 simulation are marked with different colors or backgrounds. Shading is the 95% confidence interval.

Although CMIP6-MME does not accurately reproduce historical SNMD variability (e.g., failing to capture the negative north–south correlation before 2004), it successfully captures the post-2004 strengthening of the SNMD north–south relationship (Fig. 6d), aligning well with observations. The projected increase in regional temperature uniformity opposes the SNMD structure, implying a further disappearance of the SNMD mode in the future, giving way to a more spatially uniform temperature mode (i.e., SNUW) across eastern China. Considering that Barents–Kara sea-ice is already at a very low level, future warming may lead to Barents-Kara Sea ice-free summers. Since sea-ice can no longer influence the climate in eastern China, we did not further examine the future SIC projections in CMIP6-MME. Nevertheless, Barents–Kara Sea ice in other seasons may still exert impacts on eastern China climate36,37, which merits further investigation in future work.

Discussions

Our results highlight a reversal in the interannual variability of the two leading summer T2m modes in eastern China, with the SNMD mode dominating prior to 2004, and the SNUW mode emerging as the prevailing mode thereafter. These shifts in interannual T2m variability are influenced by a combination of natural internal variability and anthropogenic warming, as corroborated by observations and model simulations. AO and PCP variabilities are associated with the increasing interannual SNUW variability, while Barents-Kara SIC is linked to the decreasing interannual SNMD variability. We suggest that the AO variability is mainly a natural internal variability while PCP variability changes are likely attributed to global warming. Furthermore, the CMIP6-MME projections indicate that escalating PCP variability will likely contribute to future increases in SNUW variability. This underscores the likelihood of more prominent EHEs, thereby posing greater challenges for managing climate extremes for eastern China under ongoing global warming. Meanwhile, the projected increase in regional temperature uniformity is expected to further suppress SNMD variability. The near disappearance of the SNMD mode also implies a strengthening of SNUW mode, as temperature fluctuations will become increasingly uniform across the entire eastern China region.

It should be noted that although the roles of Arctic sea-ice, AO and PCP in the shifting modes of eastern China T2m variability is contemporaneous, and their correlations are highest simultaneously (Supplementary Fig. 16), the causal relationship between them can be obtained from the previous literature as well as from existing analyses. For example, models are able to reproduce the effect of precipitation near Pakistan on temperature in China23 and the role of Arctic sea-ice on precipitation in China40. Besides AO and PCP, SST may also contribute to the changing SNUW/SNMD variability. Therefore, we examine the relationship between SST variability and T2m mode variability (Supplementary Fig. 17). We find that those SSTs with significant interdecadal variability changes have no significant correlation with either SNUW or SNMD, suggesting SST likely has limited contribution to T2m variability changes in eastern China.

Previous studies have primarily attributed future increases in EHEs to increasing mean temperatures6,29, which are expected to exacerbate population exposures to heat extremes30,41. These changes are closely associated with the strengthening and westward expansion of the WPSH, which enhances regional warming effects6. However, Hua et al.29 highlighted that this increase in EHEs is not solely driven by rising mean temperatures, but also by a flattening of the temperature probability distribution (increasing probability of extreme temperatures). Building on these findings, our study emphasizes the critical role of temperature variability changes in shaping EHE dynamics. We further provide a mechanistic explanation for these changes under current and future climates. This contributes to a more comprehensive understanding of temperature variability and its implications for climate change mitigation and adaptation strategies.

Both natural variability and anthropogenic warming contributed to temperature changes30,42,43. Distinguishing their relative contributions is crucial for assessing future warming trends and informing mitigation strategies. Previous study44 suggests that the CMIP6-MME can broadly represent the contributions of external anthropogenic forcing, while discrepancies between the observation and the CMIP6-MME estimates may reflect internal natural variability. In this way, contributions can be quantitatively estimated. However, this approach assumes that the models are capable of accurately simulating the spatial-temporal characteristics of the targeted modes, which may not hold for all cases in this study. The CMIP6-MME reproduces the significant upward trend in the SNUW index from 1984 to 2023 (Supplementary Fig. 18a versus Fig. 1a). The sliding standard deviation of SNUW also exhibits a rising trend of 0.45 × 10−2 per year. Based on CMIP6-MME, we estimate that approximately 38 ± 29% of this increase may be attributable to external anthropogenic forcing, with the remaining (62 ± 29%) contribution likely stemming from internal variability (Supplementary Fig. 18b). These percentages are well in line with the coefficient of AO + PCP (0.6 for AO and 0.4 for PCP), as we suggest that the AO variability change is natural variability while the PCP variability change is associated with global warming.

A similar analysis for SNMD variability, however, shows discrepancies. CMIP6-MME indicates a significant decreasing trend in the SNMD index from 1984 to 2023 (Supplementary Fig. 18c), unlike the trendless SNMD index in observation (Fig. 1a). Additionally, the CMIP6-MME north–south relationship for SNMD in Fig. 5d shows positive correlation coefficients for both historical and SSP585 simulations. These inconsistencies suggest that the models have limitations in accurately capturing SNMD mode. As a result, quantitative assessments of SNMD variability trends based on CMIP6 should be interpreted with great caution (Supplementary Fig. 18d). Given that the models fail to capture the north–south dipole relationship and the observed trends, and have larger biases in simulating SNMD pattern strength, the current CMIP6 ensemble cannot reliably represent the mechanisms governing SNMD variability. Therefore, we refrain from making any quantitative attribution conclusions regarding SNMD variability in this study.

Overall, while the CMIP6-MME provides useful insights into the potential roles of anthropogenic forcing in SNUW variability, the significant model-observation discrepancies, particularly for the SNMD mode, highlighting the need for future work using alternative methodologies, such as high-resolution regional climate models or observationally constrained simulations, to achieve more robust quantitative attribution. Comparing the simulation performance of SNMD across different resolution modes is a viable research approach, with higher-resolution modes potentially yielding superior simulation results. Nonetheless, present observational and modeling results suggest that the increased SNUW variability was mainly contributed from both natural variability and anthropogenic forcing. Moreover, anthropogenic forcing would further amplify the SNUW variability and lead to the disappearance of the SNMD mode as climate warms.

Methods

Observation data

The monthly reanalysis data used in the study are mainly from the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis version 5 (ERA5; original resolution of 0.25° × 0.25°)45. The T2m data was interpolated into the 1° × 1° resolution, and the horizontal wind field, vertical velocity and geopotential height data were interpolated into the 2.5° × 2.5° resolution. The ECMWF Reanalysis of the Twentieth Century (ERA-20C)46 and NCEP/NCAR Reanalysis 147 data were also used. Monthly mean sea-ice concentration data were obtained from Centennial in situ Observation-Based Estimates of the Variability of SST and Marine Meteorological Variables, version 2 (COBE2)48, with a resolution of 1° × 1°. Monthly mean precipitation data with a resolution of 2.5° × 2.5° were mainly obtained from the Global Precipitation Climatology Project (GPCP) version 2.349. The Global Precipitation Climatology Centre (GPCC, 1° × 1°) version 202250, the CPC Merged Analysis of Precipitation (CMAP, 2.5° × 2.5°) enhanced version51 and NOAA precipitation reconstruction over land (PREC/L, 0.5° × 0.5°)52 were also used for confirmation.

Climate indices and statistical methods

In this paper, the SNUW index is defined as the T2m averaged over the region (25°–35°N, 105°–123°E), relative to the climatology of 1984–2023, and the SNMD index is defined as the T2m difference between (22°–29°N, 113°–121°E) and (33°–38°N, 108°–125°E). These definitions ensure that the SNUW and SNMD indices maintain sufficiently high correlations with corresponding EOF time series to represent the variability characteristics of EOF modes. The EOF decomposition is applied over the region (22°–38°N, 105°–123°E) for standardized T2m. The significance of the EOF analysis is assessed using the North’s test. The HD index is defined as the 200-hPa geopotential height averaged over (30°–40°N, 95°–130°E), similar to the method of previous literature7,53. The PCP index is defined as mean precipitation over (18°–30°N, 60°–76°E). The AO index is defined as the difference in sea level pressure between 45°N and 80°N. The SIC index is defined as the mean sea-ice concentration over (74°–85°N, 20°–90°E). We also test the regions of T2m in EOF analysis and the boxes for defining the climate indices used in present study. Relevant results suggest that our conclusions are reliable and do not depend on the selection of specific regions (Supplementary Figs. 19–22). Significance of Pearson correlation and linear regression was assessed by the two-tailed Student’s t test.

Since the correlation between the PCP and AO indices is not significant (R = 0.05, p > 0.1), their effects can be considered approximately additive in a linear sense. To calculate the combined standard deviation of PCP and AO, we used the linear regression coefficients. The weights of PCP and AO were determined based on their regression relationships with both the HD and SNUW indices. Specifically, for PCP, the weight was calculated as the average of its regression coefficients with the HD and SNUW sliding standard deviation; that is, the mean of coefficient A (from the regression of PCP sliding standard deviation with HD sliding standard deviation) and coefficient B (from the regression of PCP sliding standard deviation with SNUW sliding standard deviation), expressed as (A + B)/2. The weight for AO was derived in the same manner. After calculating the initial weights, we further adjusted them so that the sum of the PCP and AO weights equals 1. Based on this approach, the final combined index was constructed as 0.4 × PCPSD + 0.6 × AOSD, where the SD denotes the 11-year sliding standard deviation. A similar method was applied when using the interannual indices. Although the derived weights were not exactly the same, they were close to those calculated based on the standard deviations, so we consistently used the weighting scheme of 0.4 × PCP + 0.6 × AO for simplicity and comparability. Given that PCP and AO are independent of each other, close to orthogonal (R = 0.05), and considering that both AO and PCP have positive relationship with SNUW/HD (Supplementary Table 3), thus their sum (i.e., PCP + AO) is indeed more appropriate to represent their combination effects than their difference (i.e., AO − PCP). Of course, there are some years in which the two cancel each other out (such as years 2009, 2011). However, the AO − PCP exhibits weaker and not significant correlations with SNUW (R = 0.66, p = 0.09) and HD variability (R = 0.58, p = 0.12) than that of the PCP + AO (R > 0.8, p < 0.05; see Fig. 2d). So we use the PCP + AO in this study. When removing the PCP and AO signals from the SNUW and SNMD indices, we separately subtracted the linear regression components associated with PCP and AO from the SNUW and SNMD time series. Specifically, we removed the products of the regression coefficients and the corresponding PCP and AO indices from the original SNUW and SNMD data.

Bayley and Hammersley’s method54 is used to estimate the effective degrees of freedom (\({N}_{\mathrm{eff}}\)) in time series analysis when calculating sliding correlation and sliding variability. The formula for calculating \({N}_{\mathrm{eff}}\) is:

$${N}_{\mathrm{eff}}=\frac{N}{1+2{\sum }_{k=1}^{N-1}\frac{N-k}{N}{\rho }_{x}(k){\rho }_{y}(k)}$$
(1)

where \(N\) is the number of samples. \({\rho }_{x}(k)\) and \({\rho }_{y}(k)\) are the lag \(k\) autocorrelation coefficients of the time series \(x\) and \(y\), respectively.

We use the wave activity flux (WAF)55 to effectively analyze the propagation characteristics of anomalously steady Rossby waves. The horizontal component of WAF is calculated as:

$$\mathrm{WAF}=\frac{{\rm{p}_{0}}}{2|{\rm{U}}|}\left[\begin{array}{l}{\rm{U}}({{{{\rm{\psi }}}^{{\prime} }}_{{\rm{x}}}}^{2}-{{\rm{\psi }}}^{{\prime} }{{{\rm{\psi }}}^{{\prime} }}_{\mathrm{xx}})+{\rm{V}}({{{\rm{\psi }}}^{{\prime} }}_{{\rm{x}}}{{{\rm{\psi }}}^{{\prime} }}_{{\rm{y}}}-{{\rm{\psi }}}^{{\prime} }{{{\rm{\psi }}}^{{\prime} }}_{\mathrm{xy}})\\ {\rm{U}}({{{\rm{\psi }}}^{{\prime} }}_{{\rm{x}}}{{{\rm{\psi }}}^{{\prime} }}_{{\rm{y}}}-{{\rm{\psi }}}^{{\prime} }{{{\rm{\psi }}}^{{\prime} }}_{\mathrm{xy}})+{\rm{V}}({{{{\rm{\psi }}}^{{\prime} }}_{{\rm{y}}}}^{2}-{{\rm{\psi }}}^{{\prime} }{{{\rm{\psi }}}^{{\prime} }}_{\mathrm{yy}})\end{array}\right]$$
(2)

where \({p}_{0}=\,(\mathrm{pressure}/1000\,\mathrm{hPa})\), \({{\rm{\psi }}}^{{\prime} }\) is the perturbation streamfunction relative to climatology, and U and V are the climatology of zonal and meridional winds, respectively. |U| is the magnitude of (U, V).

CMIP6 data and processing

The outputs from 29 CMIP6 models 56,57 are used (Supplementary Table 2) with historical experiments for 1850–2014 and SSP585 experiments for 2015–2100. And for each model, only the first ensemble member available is selected. Most are r1i1p1f1, except for HadGEM3-GC31-LL (r1i1p1f3), HadGEM3-GC31-MM (r1i1p1f3) and CESM2 (SSP585; r10i1p1f1). We use the near-surface air temperature, sea level pressure, precipitation and geopotential height data. The data of different model resolutions are interpolated into the same horizontal resolution of 2.5° × 2.5°. Definitions of SNUW, SNMD, HD, AO, and PCP indices are consistent with the observation. Seasonal climatology was constructed over the 1900–1999 period, and normalized climate indices referenced to the climatology were obtained. Near-surface air temperature (SAT) is used to replace the T2m in CMIP6 models.