Steady global surface warming from 1973 to 2022 but increased warming rate after 1990

The study examines how internal climate variability, particularly sea-surface temperature (SST) patterns such as ENSO, obscures short-term changes in the global mean surface temperature anomaly (GSTA) and hence the perceived rate of global warming. The research aims to clarify recent warming rates by filtering the influence of SST-driven internal variability and to assess whether the recent La Niña years (2020–2022) imply any slowdown. It also investigates whether the warming rate has changed over recent decades, especially since around 1990, in the context of evolving anthropogenic forcings: slowing CO2 emissions growth, rapidly increasing CH4 concentrations, and decreasing SO2 emissions. The purpose is to provide a clearer metric for tracking climate change and to compare observations with CMIP6 model expectations.

The paper builds on prior work documenting global temperature datasets (HadCRUT, GISTEMP, NOAA, Berkeley Earth) and studies on Earth's energy imbalance and ocean heat content (OHC), which suggest an accelerating heat uptake. It references approaches that regress out ENSO influences and previous detection/attribution frameworks. Prior work by the authors introduced a Green’s function-based method to filter SST pattern influences from GSTA, showing a 0.19 °C/decade warming over 1971–2020. Literature also discusses potential recent increases in warming rates and systematic biases in CMIP6 models (elevated ECS/TCR, misrepresentation of Pacific SST patterns), as well as uncertainties in aerosol radiative forcing and the climatic effects of volcanic eruptions like Mt. Pinatubo.

Datasets: Primary observational dataset is HadCRUT5 (v5.0.1.01) monthly gridded surface temperature anomalies from January 1850–December 2022. Additional datasets: GISTEMP v4, NOAA GlobalTemp v5.1, and Berkeley Earth (April 2023 update), all monthly, through December 2022. Trend removal: At each grid point, a 10-year boxcar smoothing is applied relative to an 1850–1900 baseline to remove long-term anthropogenic warming and spatial/seasonal patterns, leaving detrended anomalies representing internal variability. Endpoints are handled by mirroring data to maintain consistent weighting. Decadal-scale regional variability is not fully removed by this method. Green’s function (GF) filtering: A GF is computed with the CESM1.2.1-CAM5.3 Earth System Model from 40-year fixed-SST simulations, perturbing SSTs in 74 partially overlapping ocean patches (80° longitude × 40° latitude). The GF relates idealized local SST increases to global-mean surface air temperature (2 m) responses via changes in radiation, clouds, and water vapor. For each month, the detrended observed SST pattern is multiplied by the GF and summed over ocean-dominated grid points to estimate the modulation of GSTA due to SST variability. HadCRUT5 and GISTEMP fields are regridded to the GF resolution (2.5° latitude × 1.9° longitude). This modulation is subtracted from the raw GSTA to yield the filtered series. Analysis of warming rates: Linear regressions are performed on filtered and unfiltered annual mean series to estimate warming rates over 50-year (1973–2022), 20-year (2003–2022), and 10-year (2013–2022) periods. Sliding-window trend analyses (10–30 years) evaluate time evolution of warming rates and derive rate-of-change (°C/decade/decade). Uncertainties are 5–95% confidence intervals from the regressions. Climate model comparison: CMIP6 ScenarioMIP simulations (historical to 2014, SSP thereafter; SSP1-2.6 and SSP5-8.5 used, results insensitive through 2022) are processed identically, including GF filtering, using monthly 2 m air temperature. One ensemble member per model (33 models; 119 members total) is used, with multiple members for three large ensembles (CanESM5, MPI-ESM1-2-LR, ACCESS-ESM1-5) to illustrate internal variability. Observed combinations of mean 50-year warming rate and rate increase are compared against the CMIP6 ensemble, with model ECS/TCR characteristics considered.

• ENSO modulation: Estimated SST-related influence on annual GSTA is −0.09 °C in 2021 and −0.04 °C in 2022 (La Niña), compared to +0.16 °C during the 2016 El Niño.
• 50-year warming rate (1973–2022): HadCRUT5 filtered rate 0.19 ± 0.01 °C/decade; GISTEMP 0.19; NOAA 0.18; Berkeley Earth 0.17 °C/decade. Mean filtered rate across the four series is 0.18 ± 0.01 °C/decade. 2022 aligns with the 50-year trend.
• Recent periods (filtered): For 2003–2022, warming rates are higher than the 50-year average: HadCRUT5 0.21 ± 0.03, GISTEMP 0.23 ± 0.03, NOAA 0.22 ± 0.03, BEST 0.20 ± 0.03 °C/decade. For 2013–2022: HadCRUT5 0.23 ± 0.08, GISTEMP 0.26 ± 0.06, NOAA 0.25 ± 0.06, BEST 0.24 ± 0.07 °C/decade.
• Warming rate increase since ~1990: For HadCRUT5, a 20-year sliding window shows an increase in warming rate of 0.012 °C/decade/decade [0.008, 0.017] after filtering, with stronger signal-to-noise than unfiltered data. Other series (filtered, 20-year window): GISTEMP 0.022 ± 0.004, NOAA 0.025 ± 0.003, BEST 0.008 ± 0.003 °C/decade/decade. Mean filtered increase across the four series is 0.016 °C/decade/decade.
• Temporal character: The increase appears as a step-up around the early 1990s rather than a continuous acceleration; subsequent leveling is observed. This pattern is qualitatively consistent with NOAA OHC trends.
• CMIP6 comparison: Most CMIP6 simulations show higher 50-year warming rates than observations and lower rate increases for a given mean rate, with a positive association between model mean rates and rate increases tied to higher ECS/TCR. Observations lie outside the CMIP6 ensemble when considering the joint space of mean 50-year warming rate and rate increase; CMIP6 mean rate increase is ~0.007 °C/decade/decade (one member per model).

Filtering out SST-driven internal variability clarifies that observed global surface warming has continued at a steady multi-decadal rate (~0.18 °C/decade over 1973–2022) and that recent La Niña years do not indicate a slowdown. At the same time, the analysis reveals an elevated warming rate since around 1990, expressed as a step-up rather than a smooth acceleration, aligning with independent indications of accelerating ocean heat content. These findings sharpen the interpretation of near-term GSTA, disentangling internal variability from forced trends, and they underscore uncertainties in the balance of anthropogenic forcings (GHGs vs. aerosols). The discrepancy between observations and CMIP6 in the combined metrics of mean rate and rate increase highlights potential model biases (e.g., SST pattern representation, aerosol forcing, volcanic response) and has implications for near-term climate risk assessments. Anticipated El Niño conditions in 2023–2024 were expected to produce record-high GSTA given the elevated underlying rate, emphasizing the utility of the filtered metric for real-time climate monitoring.

The study updates and applies a Green’s function-based filtering method to major global temperature datasets, showing that global surface warming has proceeded at an approximately steady multi-decadal rate (~0.18 °C/decade) through 2022, with a discernible step-up in warming rates since around 1990. The method enhances signal-to-noise in trend estimates by accounting for SST-driven internal variability, enabling more robust assessment of recent changes. Observations, when evaluated jointly by mean warming rate and rate increase, fall outside the CMIP6 ensemble, raising concerns for the use of unadjusted CMIP6 outputs in impact and risk analyses. Future research should expand filtering frameworks to include multiple models and drivers (e.g., stratospheric temperature, volcanic eruptions, heterogeneous aerosol distributions), refine aerosol forcing constraints, and reassess model physics governing SST pattern responses and climate sensitivity.

• The filtering tool relies on a Green’s function derived from a single Earth System Model (CESM1.2.1-CAM5.3), which may bias the mapping from SST patterns to global mean temperature.
• Only SST-driven internal variability is explicitly filtered; other influences (e.g., volcanic eruptions, stratospheric processes, heterogeneous aerosol forcing patterns) are not directly included.
• The 10-year boxcar trend removal does not eliminate decadal-scale regional variability in temperature patterns.
• Observational datasets differ in coverage and treatment of data-sparse regions and may use different SST inputs, contributing to inconsistencies in estimated rate increases.
• Model–observation comparisons may be affected by CMIP6 model biases (e.g., ECS/TCR distributions, Pacific SST pattern representation) and scenario assumptions beyond 2014, though results through 2022 are reported to be scenario-insensitive.