Future changes in the frequency of temperature extremes may be underestimated in tropical and subtropical regions

The study investigates why projections of hot temperature extremes differ substantially among climate models even under the same level of global mean warming. Daily maximum temperature (TX) is a key metric for heat wave intensity and is influenced by radiation, heat transport, and surface energy exchanges. Soil moisture deficits and surface humidity processes can amplify TX variability and heat extremes. Many models are suspected to be too dry at present, potentially inflating TX variability and affecting projected heat wave frequency. The authors hypothesize that models’ skill in simulating present-day TX variability can constrain future projections of extreme heat frequency, and that land–atmosphere humidity feedbacks play a central role, particularly in tropical and subtropical regions.

Prior work shows detectable changes in temperature extremes at low levels of global warming and links uncertainties in projected extremes to land–atmosphere feedbacks. Studies have found that soil moisture deficits enhance surface temperature extremes and that using observed land heat fluxes can reduce projection uncertainties. There is evidence that some climate models overestimate TX variability and that interannual variability of summer temperatures over Europe is better captured by models with realistic variability. Other processes (dynamics, aerosols) also affect TX variability, and constraints related to precipitation or radiation may be more relevant in some regions.

The authors develop and apply an emergent constraint (EC) based on present-day TX variability to constrain future changes in the frequency of hot extremes. Indices: (1) TX98p, defined as the number of days above the daily climatological 98th percentile during the warm season (JJA for Northern Hemisphere, DJF for Southern Hemisphere, and year-round for 15°S–15°N), using a ±15-day window over 1995–2005 to estimate daily thresholds. (2) ATX, a variability metric computed as the difference between the 95th percentile and the mean of the daily TX distribution (using ±15-day pooling) for each calendar day and location during the warm season; ATX is defined as ATX_model – ATX_ref, where ATX_ref is from ERA5. The EC uses the relationship between historical ATX and future changes in TX98p. Datasets: 27 CMIP5 and 7 CMIP6 models are used (multi-member ensembles averaged per model where available). Projections use RCP4.5 (CMIP5) and SSP245 (CMIP6). Three targets are analyzed: end-of-century (2091–2100), and decades corresponding to +1.5 °C and +2 °C global warming above pre-industrial (implemented as +0.7 °C and +1.2 °C relative to the 1995–2005 baseline, respectively). Model results are normalized by each model’s mean Tas change for comparability and, for time series, rescaled to the multi-model Tas trend. Observational references and uncertainties: ERA5 provides daily TX (1995–2005). CHIRTS is used to estimate observational error (difference ERA5–CHIRTS). Model internal variability/error in ATX is estimated from the HAPPI historical atmosphere-only large ensembles (five models with ≥100 members; one outlier mean-bias model excluded for the main estimate). The uncertainty envelope for model selection at each grid point combines observational error and internal variability, reduced by the square root of the number of ensemble members for multi-member models. EC application: Models whose ATX falls within the combined uncertainty range are selected. The EC is applied regionally by smoothing ATX over land with a 5° spatial filter and selecting models grid-point-wise where the ATX–TX98p change correlation is significant; a global EC variant selects models using the globally averaged ATX. Robustness checks: Sensitivity tests to percentile thresholds (95th vs 98th), spatial smoothing scales, uncertainty ranges, and model removal (leave-one-out). Physical mechanism evaluation relates TX98p changes to changes in latent and sensible heat fluxes during hot days. Perfect-model tests use one model as pseudo-observation to confirm the EC reduces projection error. Alternative metrics (interannual TX variability, DTR variability, latent and sensible heat flux variability) are also tested to avoid selection bias and confirm EC applicability across humid regions.

- A statistically significant, predominantly negative correlation exists between baseline ATX (TX variability relative to ERA5) and future TX98p change over tropical and subtropical regions and South and East Asia: models with higher present-day hot-day variability tend to project smaller increases in hot-day frequency, and vice versa. - Applying the regional EC (selecting models matching observed ATX within uncertainty) yields larger projected increases in TX98p than the unconstrained ensemble over many low-latitude regions. Africa, South and Central Asia, and South America show particularly strong amplifications, locally exceeding a 50% increase in the number of days exceeding the historical 98th percentile compared to unconstrained projections. - The constrained projections imply that the frequency of hot extremes previously expected by the end of the century could be reached around 2060 (~40 years earlier), when normalized per degree of global warming. - The EC is robust across warming targets (+1.5 °C, +2 °C, end-of-century), though the area of significant correlation is smaller at +1.5 °C. - The physical mechanism is consistent with land–atmosphere interactions: decreases in latent heat flux and increases in sensible heat flux during hot extremes correlate with higher TX98p increases in constrained models, indicating that drying and reduced evaporative cooling amplify extreme temperatures in humid and semi-humid regions. - Mid- and high-latitudes show weaker or insignificant ATX–TX98p relationships; other processes (e.g., precipitation changes, dynamics) likely dominate there. Desert regions (North Africa, Middle East) show relationships not clearly linked to latent heat flux, suggesting potential roles for shortwave radiation, aerosols, or cloud cover. - The EC passes perfect-model tests and sensitivity analyses, and similar conclusions arise when using alternative variability-related metrics. A global EC variant (seven selected models) also shows increased hot extreme frequency, though amplification is slightly weaker than the regional EC.

The findings indicate that many climate models likely underestimate future increases in the frequency of unusually hot days across low-latitude regions due to biases in present-day TX variability linked to land–atmosphere humidity exchanges. Constraining projections with observed-consistent variability increases the projected frequency of hot extremes, highlighting the importance of soil moisture and evaporative cooling in modulating heat events. The EC framework reduces uncertainty and adjusts ensemble means toward stronger changes, especially in tropical and subtropical regions and South and East Asia. In contrast, limited or absent EC signals at mid- to high latitudes and in deserts imply that different processes govern extremes there, warranting region-specific constraints (e.g., precipitation, dynamics, radiation). The results underscore the societal relevance of improved projections for health, agriculture, and water resources by mid-century and motivate incorporating physically grounded constraints in multi-model assessments.

By using an emergent constraint based on present-day daily maximum temperature variability (ATX), the study demonstrates that projections of hot extreme frequency (TX98p) are likely underestimated in tropical and subtropical regions. Selecting models consistent with observed variability (ERA5, with uncertainty informed by CHIRTS and HAPPI) increases projected hot-day frequencies, in some regions by more than 50%, and advances timelines such that end-of-century levels could be reached around 2060. The work links these changes to land–atmosphere humidity processes (reduced latent and increased sensible heating). The approach is robust across warming targets, passes perfect-model tests, and remains broadly consistent when alternative variability metrics are used. Future research should extend constraints to additional physical processes and regions (e.g., dynamics, radiation in deserts, high-latitude processes) and evaluate heat stress metrics incorporating humidity, such as wet-bulb temperature, using analogous EC methodologies.

- Regional applicability: The ATX–TX98p emergent relationship is significant mainly in tropical and subtropical regions and South and East Asia; it is weaker or absent at mid- and high latitudes and over deserts, limiting generalizability. - Mechanistic ambiguity in arid regions: Strong signals over North Africa and the Middle East are not clearly tied to land drying; radiation and aerosol processes may dominate, reducing confidence there. - Observational uncertainty: Differences between ERA5 and CHIRTS imply observational error; ERA5 tends to have weaker TX variability than some station-based datasets, though within HAPPI-based uncertainty. - Data limitations: Daily soil moisture and some surface flux outputs are limited (especially in CMIP5), so latent/sensible heat fluxes serve as proxies; this may omit other humidity-related processes. - Model selection breadth: In some regions (e.g., central Africa), few models meet the constraint, making amplification estimates more sensitive to small sample sizes. - Methodological choices: Results depend on definitions (e.g., 95th vs 98th percentile), spatial smoothing scale, and normalization by Tas; sensitivity tests show robustness but residual sensitivity remains. - Scenario and ensemble coverage: Only CMIP5 RCP4.5 and CMIP6 SSP245 were used; other scenarios or models might influence projections and constraint performance.