Lacking oceanic-driven internal multidecadal climate variability is compensated by forced variability in model simulations

TL;DR

The paper tackles the challenge of disentangling forced versus internal multidecadal variability (MDV) in regional climate change by removing CO$_2$-congruent forcing to obtain a CO$_2$-detrended temperature field. It leverages both instrumental records (HadCRUT5) and CMIP6 historical ensembles to compare spatial covariance patterns and the land–ocean contrast, revealing that MDV in observations is primarily oceanic internal variability on $30-80$ year timescales, while many models exhibit a strong, spatially coherent forced component over land. A key finding is an emergent relationship between the origin of oceanic MDV (forced vs internal) and the land–ocean variance ratio; models with higher residual forced MDV show inflated land signals and a pronounced land-ocean contrast, whereas models with higher internal oceanic MDV align better with observations. The results imply that oceanic internal variability is underrepresented in many climate models, which may lead to underestimations of the true range of internal variability in ocean-dominated regions and suggest the need to improve ocean dynamics representations and paleoclimate constraints for better regional projections.

Abstract

Regional climate change in the $21^{st}$ century will result from the interplay between human-induced changes and internal climate variability. Competing effects from greenhouse gas warming and aerosol cooling have historically caused multidecadal forced climate variations overlapping with internal variability. Despite extensive historical observations, disentangling the contributions of internal and forced variability remains debated, largely due to the uncertain magnitude of anthropogenic aerosols. Here, we show that, after removing CO$_{2}$-congruent variability, multidecadal temperature variability in instrumental data is largely attributable to internal processes of oceanic origin. This follows from an emergent relationship, identified in historical climate model simulations, between the driver of variability in oceanic regions and the land-ocean variance ratio in the mid-latitudes. Thus, climate models with higher residual (non-CO$_{2}$) forced variability, largely linked to volcanic and anthropogenic aerosols, exhibit more spatially coherent and amplified temperature patterns over land compared to observations. In contrast, models with higher internal variability agree better with the instrumental data. Our results underscore that internal modes of ocean-driven variability may be too weak in many climate models, and that current projections may be underestimating the range of internal variability in regions with high oceanic influence.

Figures (13)

• Figure 1: Fraction of the multidecadal variance explained by the CO2-congruent variability (${R^2}_{CO_2,MD}$). a, Shown is the ensemble mean ${R^2}_{CO_2,MD}$ of the instrumental data. b, Same as a, but for the ensemble mean of the models. c, Difference between the instrumental (a) and model (b) estimates. The gray "$\times$" markers indicate that less than three quarter of the models agree on the sign of the difference.
• Figure 2: Spatial extent of correlation patterns with GMST in instrumental and model data. a, Multidecadal variance explained by GMST in the local instrumental temperature timeseries after both were CO2-detrended. b, Same as a, but showing the average of the models. c, CO2-detrended GMST timeseries are shown for the instrumental data (orange) and the models (cyan); shading indicates 66% spread of the respective ensembles. The analogous result for the internal variability only of the models is also shown (gray). d, Spatial extent of the patterns in a,b given as the fraction of the Earth's land area that was above the $r^2>0.25$ threshold; also shown for comparison is the same measure calculated from the internal variability in the models (gray distribution). e, Same as d, but for Earth's ocean area.
• Figure 3: Land-ocean constrast across the mid-latitudes. a, Power spectral density of CO2-detrended mean temperature for the North Atlantic region, comparing instrumental and model data. Shading indicates the 66% spread of the respective ensembles, and the range of multidecadal timescales (30-80 years) is highlighted by vertical dashed lines. b, Same as a, but for the Central Asia region. c, Comparison of the land-ocean multidecadal variance ratio (Central Asia over North Atlantic) between CO2-detrended instrumental and model data. d, Opposite variations in multidecadal variance are observed between the CO2-detrended instrumental and model data across longitudinally sliding boxes over the northern mid-latitudes (see Supp. Fig. \ref{['ExtDataFig_3']} for the undetrended result). Shading shows the 66% spread of the respective ensembles.
• Figure 4: Relationship between the land-ocean contrast in MDV and the relative contribution of internal vs forced oceanic variability. The correlation $r_{\sigma_{MD}^2,RLI}$ calculated from the CO2-detrended models across the northern mid-latitudes is shown as a function of the logarithm of the ratio of residual forced to internal oceanic multidecadal variance. Residual forced variability is estimated from the CO2-detrended single-model ensemble means, and internal variability from the individual realizations after the single-model ensemble means were removed. The regression line (dashed black) and corresponding correlation value are shown, highlighting the direct relationship between the driver of oceanic variability and the land-ocean contrast strength. The $r_{\sigma_{MD}^2,RLI}$ values calculate for the instrumental data are shown as a boxplot for comparison (orange boxplot, circles indicate outliers) and arbitrarily placed at the leftmost value on the x-axis; the 66% range is shown across with orange shading.
• Figure S1: Evaluation of CO2-detrending in greenhouse gas only (GHG-Only) historical experiments versus subtracting the single-model ensemble mean. a, Multidecadal variance (30-80 years) calculated from the power spectra of the CO2-detrended temperature fields, averaged across the GHG-only ensemble (comprising 90 realizations across three models: HadGEM3-GC31-LL, MIROC6 and MPI-ESM1-2-LR). b, Same as a, but with internal variability obtained by subtracting the single-model ensemble mean from each realisation. c, Same as a, but introducing data gaps matching the HadCRUT instrumental data sampling before the CO2-detrending is performed. d, Same as b, but also with sampling gaps as in c. The spatial patterns of multidecadal variance obtained by both methods are highly similar and strongly correlated ($r=0.99 \pm 0.01$ without sampling gaps, $0.97 \pm 0.01$ with sampling gaps).