Large-scale emergence of regional changes in year-to-year temperature variability by the end of the 21st century

The study addresses how human-induced climate change affects internal interannual (year-to-year) temperature variability, not just mean warming. While global mean temperature trends are relatively well constrained, changes in variability and associated extremes remain uncertain and are critically important for ecological and societal impacts. Traditional tools (instrumental records, single simulations, CMIP multi-model ensembles) struggle to disentangle forced responses from internal variability, leading to inconclusive projections ranging from no global mean change to slight decreases or regional increases in variability. Leveraging single-model initial-condition large ensembles (SMILEs), the authors aim to robustly separate internal variability from forced signals to quantify past, present, and future changes in regional temperature variability and assess when and where forced changes emerge beyond the range of unforced variability. The purpose is to establish if and how anthropogenic forcing drives a distinct, large-scale pattern of temperature variability change and to evaluate consistency with instrumental and paleoclimate evidence, thereby informing risk assessments and mitigation urgency.

Prior work indicates uncertainty in future temperature variability: studies have reported no change or slight global-mean decreases alongside notable regional increases. Difficulties arise from the non-stationary single realisation of observations and limited ability of standard ensembles to separate forced and internal variability. ENSO-related variability in the tropical Pacific and its response to warming remain uncertain due to competing feedbacks. Past climate reconstructions (e.g., PAGES2k) suggest limited sensitivity to high-frequency variability, while volcanic forcing prominently modulates variability over the last millennium. Tree-ring-based N-TREND reconstructions of Northern Hemisphere summer temperature variability broadly align with PMIP3 and CESM1-CAM5 LME patterns, despite known regional biases (e.g., Quebec) and lower observed variability amplitudes. Overall, the literature highlights the need for tools like SMILEs to robustly quantify internal variability and its changes under external forcing.

Data and models: - Paleoclimate: PAGES2k multiproxy global temperature database; N-TREND gridded NH summer tree-ring temperatures. - Instrumental/reanalyses: HadCRUT5 (non-infilled), GISTEMP, HadCRUT4-CW, ERA-20C, NOAA-20C. - Models: 10 SMILEs (7 CMIP5: CanESM2, CESM1-CAM5, CSIRO-Mk3.6.0, EC-EARTH, GFDL-CM3, GFDL-ESM2M, MPI-ESM-LR; 3 CMIP6: CanESM5, IPSL-CM6A-LR, MIROC6) with historical and high-emission scenarios (SSP5-8.5 or RCP8.5); CESM1-CAM5 Last Millennium Ensemble (LME; all-forcing and single-forcing experiments); PMIP3 last-millennium simulations. - Variables: near-surface air temperature, sea ice concentration, sensible and latent heat (to compute Bowen ratio). Preprocessing: - All datasets remapped to 1°×1° grid via bilinear interpolation. - Model last-millennium and preindustrial control simulations linearly detrended to remove drift. - Observational products detrended using the multi-model mean of eight SMILE means (1920–2019) to approximate the forced signal. - Observational masking: non-infilled HadCRUT5 used to define coverage; grid cells with <25 years of data in 1920–1969 masked; same mask applied across observational products and for model–data comparisons. Quantifying variability: - Observations: temporal standard deviation over 50-year periods (1920–1969 and 1970–2019) after detrending. - Models/SMILEs: annual internal variability estimated as sample ensemble standard deviation across ensemble members for each year (quasi-ergodic assumption consistency with time variance). Smoothed using centred running means (20-year for Fig. 1; 10-year for Fig. 3 periods). - Distinction: globally averaged regional temperature variability (grid-point SD averaged globally) vs. SD of global mean temperature. Change metrics and comparisons: - Spatial patterns: ratios of variability between periods (e.g., 1970–2019 vs 1920–1969 in observations; 2010–2019 and 2090–2099 vs preindustrial control variability in SMILEs). - Regional aggregates: tropical land (30°N–30°S), high latitudes (90–50°N and 50–90°S), and global average of regional variability. - Natural external forcing benchmark: CESM1-CAM5 LME used to estimate ranges of naturally forced changes in 10-year averaged internal variability. - Emergence definition: variability in 2080–2099 compared against full range from consecutive overlapping 100-year windows in each model’s preindustrial control. Emergence if end-century variability lies outside the min–max of control SDs (above or below). Significance testing: - Grid-point changes tested using two-sided F-test at 5% significance; note dependence on length of control for stringency. Mechanism diagnostics: - High latitudes: assessed relationship between NH sea ice area and 60–90°N mean temperature and temperature variability over 1850–2100 (models and observations for sea ice vs temperature), linking sea ice decline to reduced variability via increased ocean heat capacity and altered gradients. - Tropics/land: computed Bowen ratio (sensible/latent heat) changes (2090–2099 vs 1950–1959) and compared spatially with changes in temperature variability to infer roles of drying, vegetation change, soil moisture feedbacks; noted exceptions (e.g., parts of North/East Africa with increased precipitation/vegetation). Scenarios and periods: - Present: 2010–2019 relative to preindustrial. - Future: 2090–2099 under SSP5-8.5 or RCP8.5 relative to preindustrial. - Long-term context: 850–1849 from PMIP3 and CESM1-CAM5 LME to evaluate natural variability and volcanic impacts.

- Emergence: Anthropogenically forced changes in internal interannual temperature variability are projected to emerge from the unforced range by late 21st century, with consistent regional patterns across SMILEs despite differing global-mean responses. - Latitudinally contrasting pattern: • Present decade (2010–2019): Tropical land variability increases on average by +6.5% (range 0.8–14.8%) relative to preindustrial; high-latitude variability decreases by −6.4% (range −18.3% to 0.4%). • End-century (2090–2099; SSP5-8.5/RCP8.5): Tropical land variability increases by +12.7% (range 1.5–26.0%); high-latitude variability decreases by −23.7% (range −39.7% to 9.8%); globally averaged regional variability changes by −3.7% (range −9.1% to 4.3%). - Global mean inconsistency: Models disagree on sign and magnitude of globally averaged regional variability change; some decrease continuously, others show non-monotonic changes or no change. By 2080–2099, five SMILEs fall below and two rise above their unforced ranges; three do not emerge globally. - Spatial hotspots: Consistent increases over Amazon, Southeast Asia, Australia, and West Africa; consistent decreases over Arctic/Antarctic oceans and adjacent lands under strong warming. Oceanic tropical Pacific responses vary across models, reflecting ENSO variability/uncertainty. - Observations and proxies: Instrumental records (1970–2019 vs 1920–1969) broadly corroborate increased variability over tropical/subtropical land and central/eastern Pacific and decreases at mid-latitudes, though observed changes often stronger; data gaps limit confidence at high latitudes and some tropical regions. Paleoclimate reconstructions (PAGES2k, N-TREND) support simulated magnitudes and spatial patterns; models show higher variability than N-TREND but within inter-model spread. - Mechanisms: • High latitudes: Sea ice loss driven by Arctic amplification reduces temperature variability due to higher ocean heat capacity and reduced meridional/land-sea temperature gradients; transient increases can occur during seasonal ice-transition phases before decline. • Tropics/land: Increased variability linked to drying and vegetation changes indicated by rising Bowen ratio; land-use change, fires, precipitation shifts, and enhanced evaporation contribute. Some North/East African regions may see increased variability with increased vegetation/precipitation. - Natural forcings context: Volcanic eruptions dominate naturally forced variability over the last millennium; anthropogenic forcing produces a distinct, regionally contrasting pattern exceeding natural-forcing-induced changes, already evident in present-day tropical land variability.

The study demonstrates that human forcing imprints a robust, large-scale pattern on year-to-year temperature variability, addressing longstanding uncertainty about whether variability would rise or fall globally. While global-mean regional variability does not provide a consistent signal across models, the physically interpretable and spatially coherent response—enhanced variability over tropical land and reduced variability at high latitudes—emerges strongly and is consistent with observed trends, paleoclimate evidence, and underlying mechanisms. The findings imply that climate risk is reshaped not only by mean warming but also by variability changes: tropical regions—home to dense populations and biodiversity—face heightened interannual temperature variability and associated heat extremes, whereas high-latitude regions see damped variability as sea ice disappears. The emergence beyond the preindustrial variability envelope underscores that future variability states will be unprecedented in the instrumental and last-millennium context. Differences across models in global averages largely reflect the trade-off between the opposing regional signals and differences in the timing/magnitude of sea ice loss and tropical land-surface drying. The analysis also highlights the influence of ENSO variability and its model uncertainty over tropical oceans, and the role of initial sea ice amounts and Arctic warming rates in pacing high-latitude changes. Overall, the results reinforce the importance of considering variability alongside mean changes in adaptation and mitigation planning.

This work integrates SMILEs, instrumental records, and paleoclimate reconstructions to quantify past, present, and future changes in internal interannual temperature variability. It reveals a robust anthropogenic pattern: increasing variability over tropical land and decreasing variability at high latitudes, with forced changes emerging beyond the unforced preindustrial range by century’s end under strong warming scenarios. Mechanistic links to sea ice loss and vegetation/land-surface drying explain the contrasting regional responses. The present-day onset of these changes is broadly supported by observations, despite data limitations. The results argue for urgent climate mitigation to avoid unprecedented variability increases in the tropics and associated extremes. Potential future research directions include: improving observational coverage and homogenisation in data-sparse tropical land and high-latitude ocean regions; better constraining ENSO’s response to warming; enhancing representation of dynamic vegetation, land-use change, and soil-moisture processes in models; and extending large ensembles and long control runs to refine regional emergence and significance assessments.

- Observational constraints: Significant data gaps and uncertainties in key regions (tropical land, high-latitude oceans), especially early in the record; reliance on non-infilled HadCRUT5 mask reduces spatial coverage and likely yields lower-bound variability estimates. - Proxy limitations: PAGES2k exhibits reduced sensitivity to high-frequency variability; N-TREND has known regional biases (e.g., Quebec) and generally lower variability amplitude than models. - Model sampling and design: Limited number of SMILEs (10) and varying ensemble sizes (e.g., IPSL-CM6A-LR uses 6 members for scenarios); regional sampling may still be insufficient for all areas. Differences in vegetation modelling across SMILEs (dynamic vs prognostic vs prescribed) add structural uncertainty. - Scenario scope: Future projections focus on high-emission scenarios (SSP5-8.5/RCP8.5); variability responses under lower-forcing pathways are not quantified here. - Statistical choices: Emergence defined relative to 100-year segments from preindustrial controls; significance (F-test) sensitivity to control length; detrending observations using SMILE multi-model mean assumes adequate forced-signal estimation. - Process uncertainty: ENSO variability and its response to warming remain uncertain, contributing to low model agreement over tropical oceans; timing of observed vs simulated sea ice loss can differ, affecting high-latitude comparisons.