Main

There is growing evidence that global warming is leading to an increase in the frequency and intensity of extreme weather events1,2,3,4,5,6, based on the 27 extreme event indices recommended by the Expert Team on Climate Change Detection and Indices (ETCCDI), jointly established by the World Meteorological Organization (WMO) and the World Climate Research Programme (WCRP). However, most ETCCDI indices are defined using percentiles that describe temperature or precipitation extremes on individual days but do not directly capture rapid temperature changes between consecutive days. The day-to-day temperature change (DTDT), defined as the absolute value of DTDT difference, is a fundamental aspect of climate change and is already changing under global warming7,8,9,10,11,12. Thus, extreme DTDT events, defined as DTDTs exceeding a threshold of historical records, could be considered as candidates for these distinct extremes. This is of considerable concern given the high vulnerability of human activities11,13,14, ecosystems15 and even economic growth rates16 to DTDT. For example, DTDT can cause high public mortality and lead to various diseases by impairing the human immune system11,13,17,18,19,20,21,22,23. However, the statistical characterization and long-term evolution of such DTDT under global warming remain poorly documented, constituting a key knowledge gap.

We define extreme DTDT events and select the 90th percentile threshold on the basis of the absolute values for defining this index. First, we define extreme DTDT events of the near-surface maximum temperature as those events where the absolute temperature difference between two consecutive days exceeds the 90th percentile threshold (Extended Data Fig. 1 and Methods). The rest are defined as non-extreme events. We examine the amplitude, frequency and total intensity of these extremes following the studies of extreme heatwave events5,24. This approach provides a simple and direct way to characterize extremes on a day-to-day scale. The selection of the 90th percentile threshold is based on the severe health and ecological impacts associated with DTDT events exceeding 4–6 °C (refs. 19,25), as well as the long-term changes in the distribution of DTDT. Thus, the 90th percentile threshold serves as an appropriate criterion for identifying extreme DTDT events. Notably, both eastern China and western USA experienced record-breaking spring extreme DTDT events on 16 March 2022 and 20 May 2022, respectively (Supplementary Fig. 1), suggesting a possible increase in such extremes under climate change.

Several studies using climate models have examined changes in daily temperature variability under global warming and demonstrated that these variations are primarily driven by GHG forcing26,27. However, it remains unclear whether extreme DTDT swings, that is, the extreme temperature difference between consecutive days, have already increased, and what the global distribution has been over the past decades under global warming. More importantly, the mechanisms driving the observed changes in these extremes remain unknown. Here we present a comprehensive understanding of the observed changes in extreme DTDT over the past decades, covering facts, anthropogenic contributions, future projections and potential mechanisms. On the basis of observations28, reanalysis datasets29,30 and Earth System Model (ESM) simulation results from the Coupled Model Intercomparison Project Phase 6 (CMIP6) from 1961 to 2100, we show that the amplitude, frequency and total intensity of these events will increase in most regions at low and mid-latitudes, and the amplification in these extremes is primarily driven by the drier soil and enhanced variability in pressure and soil moisture due to anthropogenic GHG forcing. Finally, we show that the return period of the record-breaking DTDT events has decreased by a factor of tens to hundreds in the western USA and eastern China.

Extreme daily swing and ETCCDI indices

To demonstrate the significance and necessity of analysing extreme DTDT, we show that these extremes of daily maximum temperature are largely independent of 15 ETCCDI temperature-related indices across nearly all global land areas. Extended Data Fig. 2 and Supplementary Fig. 2 present the spatial distribution of the correlation coefficients between the amplitude and frequency of these extremes and the 15 ETCCDI extreme temperature indices. For these 15 ETCCDI indices, only 1–8% and 2–10% of land grid points show highly significant correlations (P < 0.01) with the amplitude and frequency of extreme DTDT events, respectively, indicating that 90–99% of grid points exhibit no highly significant association with the occurrence of these extremes. Specifically, for warm days and cool days, only 4–8% of land grid points show highly significant correlations.

Our focus on these events is further motivated by their distinct health implications11,13,19,22, which differ from those of sustained heat or cold. Moreover, our analyses reveal that the extreme DTDT is strongly associated with increased mortality, and the effects of extreme swings in maximum temperature are stronger than those in minimum temperature and diurnal temperature range (DTR) in both the contiguous USA and East China (Supplementary Figs. 3 and 4).

Among the ETCCDI indices, the DTR captures intraday temperature variability. However, there are notable differences between DTR and DTDT. The latter shows DTDT variability, manifested in two important aspects: (1) greater instability in DTDT, that is the DTDT index exhibits significantly greater instability than the DTR index (Extended Data Fig. 3) and (2) divergent responses to warming, for example previous studies31 indicate that the physical mechanisms controlling daily maximum and minimum temperatures are different. This leads to asymmetric responses to global warming. Specifically, maximum temperatures have risen less sharply than minimum temperatures, driving a pronounced decline in DTR32 (Extended Data Fig. 3a).

Observed changes

We initiate our analysis by exploring historical variations in the extreme DTDT of daily maximum temperatures at each grid cell across the global land surface during the period 1961–2020, using observational records from the Berkeley Earth project28 and reanalysis outputs from the National Centers for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR)29 and the ERA530. Thresholds of 95%, 98% and 99% were also used to ensure the robustness of these extremes. In general, extreme temperature swings show greater amplitude at mid- and high-latitudes and weaker amplitude in the tropical zone (Extended Data Fig. 4).

The decadal changes in the annual mean extreme DTDT from 1961 to 2020 are depicted for the Berkeley Earth, ERA5 and NCEP/NCAR datasets in Fig. 1 and Extended Data Figs. 5 and 6, respectively. Both datasets consistently reveal broad-scale changes in these extremes. High-latitude regions with historically high values tend to exhibit a decreasing trend, while low- and mid-latitude regions with lower values show an increasing trend over the study period. The resulting patterns of change in extreme DTDTs exhibit remarkable similarity to those previously documented for daily temperature variability26,27. Specifically, the significant increases in the amplitude, frequency and total intensity of these extreme events occurred mainly at low and mid-latitudes at rates of 0.03–0.04 °C, 0.4–0.5 °C and 4–5 °C per decade, respectively. Substantial increases in extreme DTDTs are predominantly detected over low- and mid-latitude regions in both hemispheres (Extended Data Fig. 7). The increases over the western USA, eastern China, South America and the Mediterranean region are particularly pronounced. The increases in the total intensity of these extremes over the western USA, eastern China, South America and the Mediterranean region are 11.1 °C per decade, 9.4 °C per decade, 12.4 °C per decade and 7.1 °C per decade, respectively. Furthermore, as shown in Extended Data Fig. 7, extreme DTDT in the four seasons shows very similar spatial patterns of trends and these trends are generally consistent with those of the annual mean extreme temperature variability. Therefore, the following investigations use annual mean results.

Fig. 1: Observed changes in the frequency and intensity of extreme DTDT events of the daily maximum air temperature from 1961 to 2020.
figure 1

a–c, Long-term trends in amplitude (°C per decade) (a), frequency (per decade) (b) and total intensity (°C per decade) (c) of annual extreme DTDTs over the historical period of 1961–2020. Warm (red) colours represent an increase and cold (blue) colours represent a decrease. The dotted area indicates that the trend in these extremes reaches the significance level of 0.05. The observed changes were calculated on the basis of the Berkeley Earth datasets. Basemaps from Natural Earth (https://www.naturalearthdata.com).

Full size image

Considering that the definition of extreme DTDT events does not distinguish between warming and cooling directions, we further analyse the long-term trend distributions of day-to-day warming and day-to-day cooling events. We find that their spatial patterns are largely consistent with those of the total field of these extremes, with only minor regional differences (Supplementary Figs. 5–7). The result is generally consistent with the changes in the sparse temperature flips between warm and cold extremes at a 5-day timescale33.

We also perform several sensitivity tests on the datasets and thresholds. Note that the above results are not sensitive to the selection of the NCEP/NCAR or ERA5 reanalysis datasets or the Berkeley Earth observation dataset, or to the selection of different threshold values (Fig. 1, Extended Data Figs. 5–7 and Supplementary Fig. 8). For the trend pattern, the spatial correlation coefficients of the amplitude of extreme DTDT between ERA5 and Berkeley Earth and between NCEP and Berkeley Earth are 0.51 and 0.36, respectively, and are statistically significant at the 99% confidence level. In addition, changes in these extreme events contributed to ~80% of the increase in the DTDT at low and mid-latitudes and ~70% of the decrease in the DTDT at northern high latitudes.

Attribution of historical changes

The long-term changes in the frequency and amplitude of extreme DTDT events prompt us to consider the historical attribution. We identify notable historical shifts in these extremes from CMIP6 model simulations driven by combined anthropogenic and natural forcings (ALL), as shown in Methods and Fig. 2a–c. For comparison, changes in events under individual forcing scenarios, including well-mixed GHGs, anthropogenic aerosols (AAs) and natural factors (NATs), are presented in Fig. 2d–f and Extended Data Fig. 8. Overall, the spatial patterns of changes in extreme DTDT swings and their climatological means under ALL-forcing simulations align well with those from observational and reanalysis datasets, although the simulated trends are generally weaker than those observed (Fig. 2). The pattern correlation between Berkeley Earth and the multimodel ensemble mean (MME) from CMIP6 simulations is 0.47, which is also significant at the 99% confidence level. Additionally, the distribution of trend pattern correlations between Berkeley Earth and individual models shows a consistent spatial agreement (average r = 0.46). The ALL-forced simulations successfully capture the widespread enhancement in the frequency, amplitude and total intensity of these extremes at low and mid-latitudes, along with a decreasing trend at high latitudes.

Fig. 2: Simulated changes, detection and attribution in the frequency and intensity of annual extreme DTDT events of the daily maximum air temperature.
figure 2

a–c, Long-term trends in the amplitude (°C per decade) (a), frequency (per decade) (b) and total intensity (°C per decade) (c) of annual extreme DTDT events over global land areas in CMIP6 historical simulations under ALL forcing from 1961 to 2014 (HIST-ALL). The dots indicate agreement of 70% models on the sign of change. d–f, Long-term trends in the amplitude (°C per decade) (d), frequency (per decade) (e) and total intensity (°C per decade) (f) of annual extreme DTDT events over global land areas in CMIP6 historical simulations under anthropogenic GHG-only forcing from 1961 to 2014 (HIST-GHG). g, Optimal fingerprint detection and attribution analysis (Methods) on the observed changes in the amplitude of these extremes in the Berkeley Earth dataset from 1961 to 2014. The lines denote the scaling factors of one-signal analysis for simulations under ALL forcings, of two-signal analysis for GHG and NAT, and for GHG and AA forcings, of three-signal analysis for GHG, AA and NAT forcings. Error bars indicate the best estimate of scaling factors and the 5% to 95% confidence intervals. The horizontal black solid line and grey dashed line denote the standard line of zero and unity, respectively. Basemaps in a–f from Natural Earth (https://www.naturalearthdata.com).

Full size image

Moreover, the extreme DTDT patterns simulated under GHG-only forcing closely resemble those in both the ALL-forcing simulations and the observational data (Fig. 2a–c). In particular, we can see that the GHG and ALL runs are in good agreement over the last decades in the northern high latitudes, as well as low and mid-latitudes. During the period from 1961 to 2014, the simulated increases in these extremes over the low and mid-latitudes in the GHG runs are consistent with the increases in the ALL runs (Fig. 2a–f). Over the northern high latitudes, robust decreases in regional-scale extreme DTDT are observed in both the ALL and GHG forcings. These findings indicate that the pervasive alterations in these events are mainly attributable to rising anthropogenic GHG emissions. The AA and NAT results mainly show insignificant changes (Extended Data Fig. 8).

To more precisely detect anthropogenic influences, we conduct an optimal fingerprint detection and attribution analysis (Methods). In single-signal detection, the effect of ALL forcings was successfully detected (scaling factor significantly greater than zero). Given the correlations among different forcing signals, we further apply two-signal and three-signal detection approaches to distinguish the independent contributions of individual forcings. The results demonstrate that the influence of GHGs could be clearly separated and detected from the effects of AAs and NATs (Fig. 2g and Supplementary Figs. 9 and 10). This finding provides clear evidence that the observed increase in amplitude of extreme DTDT across Berkeley Earth (Fig. 2g) and ERA5 datasets (Supplementary Fig. 10) during 1961–2014 can be attributed to anthropogenic external forcing and, more specifically, to GHG emissions. The supralinearity of the scaling factor may partly arise from insufficient resolution in the model simulations (Supplementary Figs. 11 and 12).

We then identify the potential physical reasons for the increased extreme DTDT at low and mid-latitudes under global warming. That is, anthropogenic GHG emissions have led to climate warming, which increases the synoptic sea-level pressure variability34,35,36 and soil moisture variability37 and decreases the mean soil moisture37,38. Consistent with this hypothesis, these extremes are closely linked to heightened variability in sea-level pressure, wind and soil moisture and declined mean soil moisture (Fig. 3, Supplementary Figs. 13–17 and Methods). These changes further amplify day-to-day fluctuations in precipitation39 and downwelling radiation and decline surface heat capacity37,38,40. Thus, rising sea-level pressure and soil moisture variability and decreasing soil moisture emerged as the likely thermodynamic contributors of the increase in these events in recent decades. In contrast, the decreased extreme DTDT in the northern high latitudes was probably related to the GHG-induced change in the horizontal temperature gradient caused by Arctic amplification (Supplementary Fig. 18). This is partly because northerly winds and associated cold days are warming more rapidly than southerly winds and warm days and so the Arctic amplification acts to reduce subseasonal temperature variance41.

Fig. 3: Contributions of day-to-day long-term changes in different variables to extreme DTDT variability long-term changes within 40° S–40° N.
figure 3

a–j, ERA5 reanalysis (a–e) and CMIP6 multimodel historical simulations (f–j). The contributions to long-term changes in the frequency (yr−1) of these extremes were derived using the composite change index method (Methods), which fits the spatial distribution of these extremes through a linear weighted sum of long-term changes in several variables. Variables analysed include annual DTD in sea-level pressure (DTD_slp) (a,f); annual DTD in meridional wind (DTD_vas) (b,g); annual DTD in downwelling shortwave radiation (DTD_sd) (c,h); annual DTD in soil moisture (DTD_soilm) (d,i); and annual mean in soil moisture (MEAN_soilm) (e,j). Dots in f–j indicate ≥70% model agreement on the sign of change. Basemaps from Natural Earth (https://www.naturalearthdata.com).

Full size image

Note that the extreme DTDT increased at northern high-latitude regions in AA historical runs, with significant intensification primarily occurring in the northern regions of East North America and North Eurasia (Extended Data Fig. 9). Over the past few decades, there has been a rapid reduction in AAs in North America and Europe and a rapid increase in low and mid-latitudes. These changes are evident in the AA burden and associated radiative forcings, leading to notable warming in the USA and Europe and a significant increase in the meridional temperature gradient over northern regions of East North America and North Eurasia, thereby increasing these extremes.

Projected future changes

Looking towards the future, CMIP6 ESMs predict continued widespread alterations in extreme DTDT under various Shared Socioeconomic Pathway (SSP) scenarios (Fig. 4). Under both SSP 2-4.5 and SSP 5-8.5 scenarios, these extremes are projected to intensify primarily across low- and mid-latitude regions and over broad land areas in the Southern Hemisphere, while they are expected to decline at northern high latitudes, during the long-term (2050–2099) future periods (Fig. 4), suggesting that over 80% of the population of the world will experience increased temperature extremes in the future. These projected changes are amplified as GHG concentrations rise. Consistent with the patterns shown in Fig. 4a–f, stronger responses in these events are evident under the high-emissions SSP 5-8.5 scenario compared with the moderate-emissions SSP 2-4.5 pathway. The spatially averaged frequency, amplitude and total intensity of extreme DTDT in low and mid-latitudes will continue to increase by up to ~17%, ~3% and ~20%, respectively, under the high-emissions scenario (SSP 5-8.5, Fig. 4g–i). In the future, the change in these extremes will contribute to ~80% of the increase in the amplitude and ~100% of the total intensity of total DTDT at low and mid-latitudes.

Fig. 4: Projected changes in the annual extreme DTDT for the daily maximum temperature in a warming climate.
figure 4

a–c, Projection of the difference in the amplitude (°C per decade) (a), frequency (per decade) (b) and total intensity (°C per decade) (c) of annual extreme DTDT of 11 models under an SSP 2-4.5 emission scenario from 2050 to 2099 minus that in a historical run from 1960 to 2009. d–f, Projection of the difference in the amplitude (°C per decade) (d), frequency (per decade) (e) and total intensity (°C per decade) (f) of annual extreme DTDT of 11 models under an SSP 5-8.5 emission scenario from 2050 to 2099 minus that in a historical run from 1960 to 2009. g–i, MME in the yearly time series of the amplitude (g), frequency (h) and total intensity (i) of spatially averaged extreme DTDT in tropics and subtropics (45° S–45° N) during the historical (1961–2014) and future periods (2015–2100). The simulated and projected changes were calculated on the basis of the CMIP6 outputs. The shading represents ± 1 s.d. of multimodel outputs. Basemaps in a–f from Natural Earth (https://www.naturalearthdata.com).

Full size image

Estimation of changes in the return period

Notably, both eastern China and western USA experienced record-breaking extreme spring DTDT events on 16 March 2022 and 20 May 2022, respectively (Supplementary Fig. 1). For instance, the previous record holders for extreme DTDT events over eastern China and western USA were 9 March 2013 and 22 April 2018, with temperature anomalies of DTDT at 20.4 °C and 20.0 °C, respectively. The corresponding values for 2022 are as high as 22.9 °C and 20.3 °C and this amounts to an offset of 3.16 and 3.18 standard deviations from the mean. Such extreme values pose an urgent need to calculate the change in the return period for the extreme DTDT events. Thus, we test the change in the return period for two record-breaking events in 2022 over eastern China and western USA (Fig. 5). These two record-breaking events occurred only once every 1,000–3,000 years between 1950 and 1985 but increased to once every 40–60 years between 1986 and 2021. This significant change in frequency clearly indicates a rapid increase in the frequency of occurrence of this extreme event, potentially impacting human health (Methods).

Fig. 5: Return period change in record-breaking spring extreme DTDT events in East China and the continental USA.
figure 5

a,b, Return periods fitted by using the generalized extreme value distribution (curve, with 5–95% uncertainty range shaded) of the maximum value of spring extreme DTDT event for a grid over East China (116.2° E, 33.3° N, Bozhou city) (a) and a grid over the USA (104.9° W, 39.6° N, Denver city) (yr) (b). The blue and red solid lines in a and b represent historical periods 1950–1985 and 1986–2021, respectively. The black dotted lines in a and b represent a record-breaking spring extreme DTDT event for East China and the USA in 2022.

Full size image

Discussion

In this research, we examine both the historical evolution and future projections of extreme DTDT over global terrestrial regions. On the basis of observations, reanalysis datasets and CMIP6 model simulations, we found that most low- and mid-latitude regions have experienced robust and widespread changes in these extremes under past anthropogenic warming and similar spatial distributions can be found for both day-to-day warming and day-to-day cooling events. The change in extreme temperature variability contributed 70–80% to the change in DTDT. Under global warming, the amplitude, frequency and total intensity of these events show pronounced and widespread increases in most low- and mid-latitude regions, but decreases in the northern high latitudes. On the basis of the optimal fingerprinting analysis, the long-term trends are predominantly driven by increasing GHG forcing, and these extremes are projected to intensify with warming until the end of the twenty-first century, especially under higher emission scenarios. We further show that the increase in these extremes in low and mid-latitudes is mainly caused by the drier soil and intensified variability in sea-level pressure and soil moisture, which lead to increased day-to-day variability in cloud and downwelling radiation. The decrease of these events in the northern high latitudes is related to the weakened meridional temperature gradient.

Our study highlights that the extreme DTDT represents a a distinct event type of weather extreme under global warming. The observed, simulated and projected amplification of these extremes reflects that most terrestrial populations and ecosystem functions will experience intensified temperature swings. Analysis of return period changes and associated mortality in the western USA and eastern China supports rapid amplification. Two record-breaking events in the western USA and eastern China occurred only once every 1,000–3,000 years between 1950 and 1985 but increased to once every 40–60 years between 1986 and 2021, causing larger event-related mortality. Extreme DTDTs exhibit near-exponential increases in mortality risk, with stronger impacts than those from minimum temperature or DTR. These extremes could trigger cascading impacts on human society, with threats to public health11,42, ecosystem functions15, agricultural security43 and economic development16. Considering that the health and ecological impacts of day-to-day warming and day-to-day cooling events within these extremes may differ13,42, these differences highlight the need for future studies to distinguish the two phases of extreme DTDTs. In addition, abrupt DTDTs severely challenge the existing capabilities of modern weather forecasting and climate prediction systems, such that improving the ability to accurately predict, adapt to and mitigate these extremes is urgently needed44.

Methods

Observed and reanalysis datasets

We use gridded observations of daily maximum and minimum temperatures at 2 m from the Berkeley Earth dataset27 with a 1° × 1° latitude/longitude grid. The temporal span of this dataset used in this study is 1961–2022. We also use daily maximum and minimum temperatures from the NCEP/NCAR reanalysis28 and ERA5 reanalysis29 covering the period 1950–2022. The NCEP/NCAR dataset is a spatially gridded dataset that has a fixed zonal resolution of 1.875° longitude and a varying Gaussian-shaped meridional resolution with an average close to 1.875° latitude, whereas the ERA5 dataset is also a spatially gridded one with a 2° × 2° latitude/longitude grid.

Statistical methods

In this study, we define the ‘secular trend’ as the linear trend over the study period, derived using least-squares regression. To assess statistical significance, we apply linear trend analysis using least-squares regression and evaluate significance on the basis of two-tailed Student’s t-tests at the 5% significance level. We use spatial correlation coefficients to evaluate the consistency of trend patterns among different reanalysis datasets, observational datasets and climate model simulations. Specifically, we apply pattern correlation metrics such as the Pearson correlation coefficient to quantify spatial agreement.

CMIP6 model outputs

We also analyse the changes in simulated extreme DTDT and non-extreme events via 11 ESM (Extended Data Table 1) in CMIP646. For detection and attribution, we use the 11 ESM outputs under four forcings: well-mixed GHG-only, AA only, NAT only and ALL forcings during the historical period earlier than 2014, since the ALL runs in most CMIP6 ESMs ending in 2014. For the future period (2015–2100), we focus on two tier-1 SSP scenarios, SSP 2-4.5 and SSP 5-8.5, which represent the intermediate and the most extreme emission scenarios, respectively. We further validate the mechanisms underlying these extremes using CanESM5 and MPI-ESM1-2-LR. All outputs of daily maximum and minimum temperatures at 2 m, daily surface sensible heat flux, precipitation, downwelling shortwave and longwave radiation, cloud cover and soil moisture from the CMIP6 ESM simulations are regridded to a resolution of 2° × 2° via bilinear interpolation.

Extreme DTDT index

In this study, we define extreme DTDT events as events in which the absolute value of the difference in the daily maximum temperature (Tmax) between two consecutive days exceeds the 90th percentile threshold. When the absolute difference between Tmax of day i and day i − 1 exceeds the 90th percentile threshold, only day i is labelled as an extreme DTDT day. We chose the daily maximum temperature (Tmax) for defining these extremes because Tmax exhibits stronger day-to-day fluctuations than the daily mean temperature. While DTDT itself is a continuous variable, these high-amplitude events are defined categorically for statistical characterization. The frequency index was defined as the number of such events in a season or a whole year. The non-extreme events were defined as DTDT events that do not qualify as extreme DTDT events. Although the event indicator is categorical, the amplitude index refers to the average magnitude of temperature changes on extreme DTDT days, expressed in degrees Celsius. The total intensity index was defined as the sum magnitudes of temperature changes on these events in a season or a whole year.

Here we define these extreme events on the basis of the following criteria: the absolute DTDT must exceed the 90th percentile of the absolute DTDT distribution of 1961–2020 (>90th percentile). Thresholds of 95%, 98% and 99% were also used to guarantee the robustness of this definition. The selection of the 90th percentile threshold for defining this extreme temperature index is based on the severe health and ecological impacts associated with DTDT events exceeding 4–6 °C (refs. 13,43). The calculation of seasonal extreme DTDT was based on the threshold value of the annual extreme events.

Figure 1 shows the results for the 90%-threshold extreme events and Extended Data Fig. 5 highlights the results of the 95%-, 98%- and 99%-threshold events for comparison. For most of the global land areas, the patterns of the spatial distribution of the climatological mean and the long-term trend of these extremes are highly stable across the different thresholds used in the analysis. In this study, the seasonal mean or annual mean extreme temperature variability is calculated by averaging the absolute values of these events in the same season and year. For each grid point, the change in extreme DTDT variability is calculated as the percentage change relative to the mean extreme DTDT value averaged over the baseline period of 1961–2010 at the same location, thus having a unit of percentage (Fig. 4). The long-term changes in percentile-based extreme indices may exhibit in-base/out-of-base inhomogeneity47. However, we note that such inhomogeneity does not affect the conclusions of this study, as the differences in long-term trends of amplitude, frequency and total intensity of extreme events between the periods 1961–2100 and 2021–2100 are negligible.

Extreme temperature change and ETCCDI indices

The ETCCDI has developed a set of standardized climate indices to monitor and analyse changes in extreme weather and climate events. These indices are widely used in climate research to assess trends in temperature and precipitation extremes24,48. The 16 temperature-related ETCCDI indices include frost days, summer days, ice days, tropical nights, maximum Tmax events, maximum Tmin events, minimum Tmax events, minimum Tmin events, cool nights, cool days, warm nights, warm days, warm spell events, cold spell events, diurnal temperature range and growing season length. Supplementary Table 1 lists the 16 temperature-related ETCCDI indices along with their definitions.

In this study, we reveal that the extreme DTDT of the daily maximum temperature is independent of other commonly used ETCCDI indices. Our analysis included: (1) spatial mapping of correlation coefficients between the annual extreme amplitude/frequency of DTDTs and each of the 15 temperature-related ETCCDI extreme indices (1961–2020); (2) quantification of statistically highly significant correlations (P < 0.01) across global land areas.

Optimal fingerprinting method

We use an optimal fingerprinting method to detect and attribute the observed changes in the amplitude of seasonal diurnal temperature difference (SDTD). This approach is based on a generalized multivariate linear regression model49,50 and can be expressed as: \({\bf{y}}={\bf{{\upbeta}}} {\rm{{X}}}+\varepsilon\). The observed changes (y) are represented as the sum of scaled responses to external forcings (X, referred to as the ‘fingerprints’) and internal climate noise (\(\varepsilon\)). The forced response X is estimated using the MME (ALL, GHG, AA and NAT) derived from climate model simulations. The regression coefficients (scaling factor) β represent the model-simulated responses to match observations. The noise term (\(\varepsilon\)), represents internal climate variability. The regularized optimal fingerprinting algorithm is used51. The regression was resolved by using the total least-squares method49. A forcing is considered detectable when its scaling factor β exhibits a 5–95% confidence interval significantly greater than zero. Scaling factors β >1 indicate that the model underestimates the observed response, whereas values <1 indicate overestimation.

Our analysis focuses on changes in the regional mean amplitude of SDTD during the period 1961–2014. Observational time series were primarily based on Berkeley Earth data, with sensitivity tests using ERA5 also evaluated. To reduce interannual noise and enhance the signal-to-noise ratio, all time series were processed using a non-overlapping 5-year mean (with the final period, 2011–2014, being a 4-year average).

The noise term (\(\varepsilon\)), representing internal climate variability, was estimated from two independent sources: (1) pre-industrial control simulations (piControl), which were divided into non-overlapping segments matching the length of the observational record, yielding 74 chunks and (2) the intra-ensemble spread of single-forcing simulations, calculated by subtracting the ensemble mean from each realization. Only models with at least two ensemble members were retained, resulting in 218 additional samples (Supplementary Table 1). A total of 292 internal variability samples were obtained, half of which were used for regression optimization and the other half for residual consistency testing.

Detection analyses began with single-signal regressions, wherein the observed anomalies were regressed separately onto model-simulated responses to four categories of external forcings: ALLs, NATs, GHGs and AAs. To account for potential correlations among forcings, we further conduct two-signal and three-signal analyses, enabling clearer separation of individual contributions. A residual consistency test50 is used to assess whether the residuals from the regression can be explained by the model-simulated internal variability, ensuring the robustness of detection results.

Underlying mechanisms of long-term variations

Following the definition of DTDT, we compute the day-to-day variability (DTD_VAR) of various variables as the annual mean of their absolute day-to-day differences. Other studies have used this metric to examine the mechanisms of DTDT variability52. Similarly, MEAN_VAR denotes the annual mean of each variable. Historical simulation shows that the sea-level pressure and soil moisture variability are highly correlated with extreme DTDTs (Supplementary Fig. 13). Furthermore, both sea-level pressure and soil moisture variability have increased in low and mid-latitudes (Supplementary Fig. 14).

We compute the spatial correlation between long-term changes in extreme day-to-day temperature variability and long-term changes in the DTD of various variables within 40° S–40° N across reanalysis datasets and CMIP6 models (Supplementary Fig. 15). We found that most patterns show a positive correlation, indicating consistency between the changes in these key variables and these extremes. However, we found significant discrepancies in spatial patterns, and the enhanced extreme temperature variability at low and mid-latitudes is probably attributable to the integrated influence of these drivers.

Here we develop a composite change index using these contributing factors. We first calculate the long-term change at each grid point for the target variable (annual extreme DTDT variability) and for each explanatory variable over a predefined time period. To identify the optimal linear combination of explanatory variables that best explains the spatial pattern of the long-term change in extreme DTDT variability (Ytarget), we perform a multiple linear regression:

$${Y}_{{\rm{target}}}(x,y)=\displaystyle \mathop{\sum }\limits_{i=1}^{N}{\beta }_{i} {X}_{i,{\rm{change}}}(x,y)+\epsilon (x,y)$$
(1)

where \({Y}_{{\rm{target}}}(x,y)\) is the spatial structure in the long-term change of the target variable, \({X}_{i,{\rm{change}}}(x,y)\) is the spatial structure in the long-term of the ith predictor, \({\beta }_{i}\) is its regression coefficient (weight) and \(\epsilon (x,y)\) is the residual. The contribution of each explanatory variable is\({{\beta }_{i} X}_{i,{\rm{trend}}}(x,y)\) and the weighted composite index (Yfitted) is defined as the sum of contributions of each explanatory variable. The explanatory variables include annual DTD in downward longwave radiation (DTD_ld), downward shortwave radiation (DTD_sd), soil moisture (DTD_soilm), precipitation (DTD_pr), cloud cover (DTD_cloud), zonal wind (DTD_uas), meridional winds (DTD_vas), sensible heat flux (DTD_hfss), latent heat flux (DTD_hfls) and sea-level pressure (DTD_slp). We further include annual MEAN in soil moisture (MEAN_soilm). We found that the correlation coefficient between every two explanatory variables was below 0.8, indicating no strong multicollinearity. We evaluate the effectiveness of the fitted field (Yfitted) in capturing the observed spatial pattern (Ytarget) using the Pearson correlation coefficient.

First, the combined long-term changes of these different explanatory variables can show a high predictive power in explaining the spatial patterns of change in extreme DTDT variability (Supplementary Figs. 15 and 16). The spatial correlation coefficient between the composite change index from ERA5 and changes in these extremes reaches 0.65, while for different CMIP6 models, it ranges between 0.5 and 0.8. This indicates that the composite change index reliably represents the spatial pattern of extreme DTDT variability changes, allowing us to assess the contributions of different variables to the long-term change in these events based on their weights.

We then examine the contributions of individual factors across different regions. Our results demonstrate that in the tropics and southern subtropics (40° S–20° N), both DTD_slp and DTD_vas exhibit positive contributions (Fig. 3). These factors enhance thermal advection and subsequently influence variability in precipitation, cloud cover, radiation and soil moisture (Supplementary Fig. 17). This chain of physical processes ultimately leads to an increased frequency of these extreme events.

This may be attributed to global warming intensifying tropical convective activity and moist static energy34,35,36, which probably favours stronger synoptic-scale disturbances and enhances pressure/wind variability. In the Northern Hemisphere subtropics (20° N–40° N), DTD_sd, DTD_soilm and MEAN_soilm show a positive contribution (Fig. 3d,e,i,j) while DTD_slp and DTD_vas show a negative contribution (Fig. 3a,b,f,g). These findings align with previous studies suggesting that warming-induced soil drying reduces surface heat capacity37,38, thereby increasing temperature variability and the occurrence of these extremes40. Additionally, warming-driven increases in soil moisture variability38 further amplify temperature variability and the frequency of extreme DTDT events. Therefore, this is a complex physical process influenced by several factors.

Estimation of return periods

Return periods of extreme DTDT events were estimated separately for observations (Berkeley Earth dataset) and CMIP6 model simulations under ALL and NAT forcings. We first identify the threshold for these extreme events based on the 90th percentile over a baseline period (for example, 1961–1990). A generalized extreme value distribution was then fitted to the seasonal maxima of these extremes using the maximum likelihood method. Return periods were calculated as the inverse of the exceedance probability of a given magnitude. All estimates were computed on a grid-cell basis and then aggregated for zonal and regional analyses.

Mortality data

The mortality dataset collected in this study includes the following: (1) the number of people who died from non-accidental causes, cardiovascular diseases and respiratory diseases in the USA every day from January 1987 to December 2000, which is derived from the Internet-Based Health and Air Pollution Surveillance System of John Hopkins University and contains 29 cities53,54, and (2) the number of people who died from non-accidental causes, cardiovascular diseases and respiratory diseases in Jiangsu Province every day from January 2016 to December 2000, which is derived from the Center of Disease Control in Jiangsu Province and contains 41 cities. The mortality rate is the number of deaths divided by the local population in each city.

Mortality from DTDT

In this study, we use a distributed lag nonlinear model55 combined with generalized additive mixed models56,57 to estimate the associations of extreme DTDTs in the maximum temperature, extreme DTDTs in the minimum temperature, DTR and Tmax, with various mortalities. To quantify the total contribution, independent effects and relative importance of the meteorological factors, we include each temperature variability variables, Tmax and the natural logarithm of mortality rates in the same model. To reduce collinearity, we use cross-basis terms rather than raw variables. We model exposure–response associations (meteorological factors versus percentage change in mortality) via a natural cubic spline with 3 d.f. and model the lag‒response association via a natural cubic spline with an intercept and 3 d.f. with a maximum lag of 30 days.

The changes in mortality related to extreme DTDT swings in these two densely populated regions (continental US states and eastern China) were also analysed. We construct the complex nonlinear and time-lagged associations of these extremes with public mortality using a generalized additive mixed model controlling for potentially measured meteorological factors in the continental US states and eastern China. This generalized additive mixed model has a high ability to simulate the spatiotemporal evolution of total mortality, cardiovascular mortality and respiratory mortality (Supplementary Fig. 3). The associations between extreme DTDTs and mortality were consistent across two densely populated regions and three categories of mortality. In the range below the threshold for these events, the associations with mortality remained relatively stable. However, for these events above the threshold, we observe an approximately exponential growth in the associations between these extremes and mortality (Supplementary Fig. 4), with high extreme DTDT variability significantly associated with markedly increased total mortality, respiratory mortality and cardiovascular mortality. This finding is consistent with previous research22. The effects of extreme DTDTs in maximum temperature were significantly greater than those in minimum temperature and DTR in both the contiguous USA and East China and greater than those in temperature alone in the contiguous USA (Supplementary Fig. 4). We also examine the differential impacts of warming and cooling events on various types of mortality (Supplementary Fig. 4). The results suggest an asymmetric response: strong warming events tend to have a more pronounced effect on mortality than cooling events, although the latter can significantly increase mortality in some regions.