Increased risk of near term global warming due to a recent AMOC weakening

The study addresses whether highly sensitive CMIP6 models, such as IPSL-CM6A-LR with high ECS and TCR, are consistent with observed historical warming and how to interpret their projections given internal climate variability. Background concerns include that several CMIP6 models exhibit ECS and TCR values exceeding IPCC AR5 likely ranges, projecting greater 21st-century warming than earlier generations and seemingly conflicting with observation-based estimates of historical sensitivity and TCR. The authors hypothesize that low-frequency internal variability, particularly linked to the Atlantic Meridional Overturning Circulation (AMOC), may have masked part of anthropogenic warming over recent decades, affecting comparisons between models and observations and estimates of historical sensitivity. The purpose is to quantify the role of internal variability—especially AMOC-driven variability—in shaping historical GSAT trends, reconcile high-sensitivity model behavior with observations, and assess implications for near-term warming risks.

The paper reviews that one-third of CMIP6 models have effective ECS above the IPCC AR5 likely range and one-fifth have TCR above the likely range, leading to higher projected warming. Prior studies attempted to constrain CMIP6 projections using recent observed warming trends; however, such constraints may inadequately account for internal variability, which may be stronger in some CMIP6 models than in CMIP5. Observation-based estimates of historical sensitivity (S_hist) and TCR_hist have large uncertainties, and variability-based methods do not rule out high ECS values. Internal variability at decadal to multicentennial scales can modulate forced trends, complicating model–observation comparisons. Reconstructions suggest AMOC weakening since the 1940s (e.g., Caesar et al.), though debated; other fingerprints and indices support a weakening. Previous work links AMOC variability to GSAT variations and highlights internal variability’s role in early 20th-century warming and multidecadal GSAT variability. Studies also discuss the pattern effect and its potential internal variability contribution, and mixed evidence on whether models under- or overestimate low-frequency variability, including from paleoclimate reconstructions.

- Model ensemble: Used 32-member IPSL-CM6A-LR ensemble of extended historical simulations (IPSL-EHS), following CMIP6 historical forcings (1850–2014), extended to 2060 with SSP2-4.5 forcings (ozone held at 2014 climatology). Members initialized from different years of a long piControl to sample multi-centennial variability. - Climate sensitivity metrics: Computed S_hist and TCR_hist per member via an energy-budget approach comparing preindustrial (1850–1879) to present-day (1999–2018). Used F2xCO2 = 3.77 W m−2 (range 3.50–3.94 for uncertainty), net ERF ΔF = 2.09 W m−2 (from RFMIP piClim-histall), ΔT as GSAT change, and ΔQ as Earth system heat uptake approximated from modeled ocean heat content increase divided by 0.89. Small piControl drifts in OHC and GSAT were removed prior to estimation. - Observations: GSAT proxy from infilled HadCRUT4-CW adjusted by 1.06; comparison with Berkeley Earth; North Atlantic SST from ERSSTv5; upper (0–700 m) OHC from IAP; used ensembles/uncertainties where applicable. - AMOC metrics and fingerprints: AMOC index defined as the annual maximum Atlantic meridional streamfunction at 20°–50°N. AMOC fingerprints include: (1) Caesar SST-based index (SPG relative to global SST, Nov–May); (2) Atlantic Multidecadal Variability (AMV) index (North Atlantic 0–60°N SST with forced signal removed using 60°S–60°N mean); (3) AOHC index: OHC(0–700 m) offshore Newfoundland minus OHC over the Subpolar Gyre; (4) AITA: interhemispheric near-surface air temperature difference (NH–SH). Low-pass filtering used (Lanczos, typical cutoff 11 years; 5 years for spectra). - Forced vs internal decomposition: Forced response approximated by ensemble mean; internal variability per member obtained by subtracting the ensemble mean. Linear models fitted to estimate contributions of forced and internal components to AMOC variations (1940–2016). - Trend and correlation analyses: Computed GSAT and AMOC trends over 1940–2016; evaluated relationships (least-squares regression, r^2). Selected a subset of members matching observed GSAT and AMOC trends and with lowest GSAT RMSE. Spatial trend comparisons for North Atlantic SST and OHC patterns. - Multi-model analysis: Assessed AMOC–GSAT 60-year trend relationships and AMOC spectra in first 500 years of piControl simulations across 13 CMIP6 models. Analyzed additional historical ensembles (CNRM-CM6-1, MPI-ESM1.1; CanESM5 screened out due to inconsistency). Significance tests: two-tailed Student t-tests (p<0.1) for historical regressions; random-phase resampling (1000 surrogates, p<0.1) for overlapping trend windows and time series correlations. - Filtering and diagnostics: Lanczos filters applied for low-frequency analysis; anomalies referenced to appropriate periods; spatial correlations for pattern comparisons.

- IPSL-CM6A-LR ensemble mean over-warms relative to observations, but some members reproduce historical GSAT trends well (e.g., member #14 with RMSE 0.14 K for 1900–2018 GSAT). - Historical sensitivity spread: S_hist and TCR_hist exhibit large spread across members (±2σ ≈ 1.25 K), indicating strong influence of internal variability; model’s ECS (4.5 K) and TCR (2.4 K) exceed ensemble-average S_hist and TCR_hist, consistent with nonlinear feedbacks and pattern effects. - Strong AMOC–GSAT linkage: Across IPSL-EHS, GSAT and AMOC trends (1940–2016) are strongly positively related (r^2 = 0.82). Members with weakest GSAT warming show strongest AMOC weakening; those matching observed GSAT and AMOC trends have negative AMOC trends. - Subset consistency: Six-member subset (#14, #4, #5, #25, #29, #30) selected by GSAT/AMOC trend agreement and low GSAT RMSE displays AMOC strengthening up to ~1940s then decline, consistent with reconstructions; ensemble mean shows weak forced AMOC changes. - AMOC fingerprints agreement: In member #14, observed and modeled trends align: Caesar index decrease −0.13 K decade−1 (observed −0.11 K decade−1); AOHC decrease −0.42×10^9 J m−2 decade−1 (observed −0.45×10^9). North Atlantic SST trend (1940–2018) 0.06 K decade−1 vs observed 0.065; spatial correlation 0.73. Subset members have North Atlantic SST/OHC patterns closest to observations. - Internal variability dominance historically: Analyses indicate internal variability primarily drove AMOC weakening since ~1940s in IPSL-EHS, with external forcing influence increasing and dominating AMOC decline into the 21st century. - Multi-model support: 10 of 13 CMIP6 models display significant positive relationships between 60-year AMOC and GSAT trends in piControl and exhibit notable multi-centennial AMOC variability; models with stronger linkages show larger GSAT variance. In large historical ensembles, CNRM-CM6 reproduces IPSL-like behavior; MPI-ESM shows a positive relationship but fingerprint mismatches; CanESM5 members over-warm relative to observations. - Near-term warming implications: The identified subset, after a historical internally driven AMOC weakening (1951–1990), experiences an internally generated AMOC strengthening in coming decades, associated with slightly higher GSAT warming rates (~0.36 K decade−1) than the ensemble mean (~0.34 K decade−1). Results imply some anthropogenic warming was masked and is likely to materialize, increasing the risk of crossing 2 °C sooner.

Findings support the hypothesis that an internally driven AMOC weakening since the mid-20th century reduced observed GSAT warming, reconciling high-sensitivity model behavior with observed trends when internal variability is accounted for. The strong linkage between AMOC trends and GSAT trends highlights AMOC’s central role in multidecadal global temperature variability. Agreement across multiple AMOC fingerprints strengthens the interpretation that internal Atlantic variability masked part of anthropogenic warming. Consequently, observationally derived S_hist and TCR_hist estimates over recent decades may be biased low. Emergent constraint approaches must explicitly incorporate low-frequency internal variability and consider individual realizations rather than ensemble means when weighting models. Near-term projections suggest slightly accelerated warming in members consistent with observed AMOC behavior, elevating the probability of exceeding the 2 °C threshold. More broadly, internal variability (including pattern effects) can modulate feedbacks and apparent sensitivity on multidecadal timescales, with implications for risk assessment and mitigation planning.

The study demonstrates that multi-centennial internal variability, particularly AMOC-driven, likely weakened AMOC since the 1940s and masked a portion of anthropogenic global warming, explaining why some high-sensitivity CMIP6 models can still be consistent with observed historical warming. Accounting for this effect implies higher near-term warming rates and increased risk of crossing 2 °C. Key contributions include quantifying the AMOC–GSAT trend linkage, validating AMOC fingerprints against observations, and showing multi-model support for the mechanism. Future research should: (i) refine observational constraints by explicitly incorporating low-frequency variability; (ii) assess the realism of model low-frequency variability using paleoclimate reconstructions (e.g., PAGES2k) and pseudo-proxy frameworks; (iii) quantify how North Atlantic variability (and its teleconnections with the Pacific) affects feedbacks and pattern effects; and (iv) improve attribution of forced versus internal AMOC changes in the historical and future periods.

- AMOC reconstruction and fingerprint uncertainties persist; Caesar index robustness is debated, and different fingerprints may reflect mixed internal/forced signals. - The IPSL-CM6A-LR model may overestimate low-frequency internal variability; realism at multicentennial scales is difficult to assess given short instrumental records and proxy uncertainties. - Forcing uncertainties (especially aerosols, non-CO2 GHGs) affect comparisons; the extended simulations hold ozone fixed at 2014 climatology, differing from official CMIP6 projections. - Energy-budget estimates (S_hist, TCR_hist) are sensitive to ΔF and heat uptake assumptions; ocean heat content dominates but non-ocean components are approximated. - Multi-model ensemble availability is limited; some models (e.g., CanESM5) do not reproduce observed historical warming compatibility, complicating generalization.