Strong temperature gradients in the ice age North Atlantic Ocean revealed by plankton biogeography

The study addresses how well climate models simulate the Last Glacial Maximum (LGM; 23–19 ka), a climate state markedly different from today with low atmospheric CO2 (~185 ppm), extensive northern hemisphere ice sheets, and altered ocean circulation. Traditional model-data evaluations rely on indirect temperature proxies that can be confounded by uncertainties in seasonality, depth attribution, and nuisance parameters. The authors propose an alternative, ecologically grounded evaluation based on the macroecological principle that community similarity declines with increasing environmental (thermal) distance. Because temperature is the dominant global driver of planktonic foraminifera assemblage composition and their thermal niches have been stable across glacial–interglacial cycles, comparing similarity–temperature decay patterns between modern core-top data and LGM assemblages provides a way to test simulated LGM temperature fields without reconstructing temperatures directly. The purpose is to diagnose whether models capture spatial gradients of glacial cooling, which are critical for climate dynamics and ecosystems, beyond globally averaged changes.

Early global LGM temperature reconstructions used transfer functions relating microfossil assemblages to seawater temperature (e.g., CLIMAP; MARGO). Although new geochemical proxies exist, assemblage-based approaches remain important. However, proxy-based reconstructions are challenged by ambiguities in depth/season attribution and confounding variables. Macroecology provides a complementary lens: community similarity typically decays with environmental distance across taxa and systems, with temperature a primary driver for marine plankton and especially planktonic foraminifera. Prior work shows foraminifera thermal niches are stable over late Quaternary cycles, supporting space-for-environment relationships as a basis for inference. Past model–data comparisons (PMIP3/PMIP4) agree on global mean cooling but disagree on regional patterns, particularly North Atlantic gradients. Evidence also points to significant roles for AMOC changes in shaping LGM ocean temperature patterns.

Data and biogeography: The authors assembled 2,085 LGM planktonic foraminifera morphospecies assemblages from 647 unique sites, increasing coverage by ~50% over MARGO, and used 3,916 unique modern core-top samples from the ForCenS database. Taxonomy was harmonized; key subspecies were lumped where necessary. They assessed compositional analogues between LGM and modern assemblages via square chord distance; 98% of LGM sites had good modern analogues, indicating niche stability. Similarity–temperature decay: For modern core-top data, Bray–Curtis similarity between pairs of assemblages was plotted against environmental temperature differences (annual mean at 50 m), showing a strong monotonic decline. For LGM, they combined fossil assemblage similarities with simulated LGM temperatures from PMIP3/PMIP4 equilibrium simulations (10 runs from 8 models; thetao fields remapped to 1° grid, using 50 m temperatures) to generate similarity–temperature decay plots and compared against the modern reference, including binning and heatmaps. They examined robustness to chronological uncertainty, spatial sampling differences, and model bias using PI control runs. Temperature reconstruction (transfer functions): To further diagnose model–data mismatches, they reconstructed LGM temperatures from assemblages using the Modern Analogue Technique (MAT), with region-specific calibrations to minimize cryptic endemism effects. Temperature estimates were similarity-weighted means of the ten best analogues (squared chord distance metric). Seasonal/vertical attribution ambiguity was resolved by evaluating which season and depth best explain variance in assemblages using CCA and variance-explained analyses across four seasons and annual mean at 14 depth levels (upper 500 m) in five regions, separating tropics vs extratropics. Globally, annual mean temperature at 50 m explained the most variance; thus, reconstructions used 50 m annual mean. Temperature anomalies were referenced to WOA98 climatology (100-km radius averaging around sites) to avoid bias from recent warming. Uncertainty and sensitivity: Reconstruction uncertainty was quantified via Monte Carlo by generating 1,000 spatially autocorrelated residual fields per region; median 2σ reconstruction error was 1.52 °C, with spatial autocorrelation accounted for. Sensitivity to poor analogues was assessed by filtering LGM samples above poor-analogue thresholds globally and regionally; effects on mean anomalies were negligible. Spatial sampling sensitivity of modern similarity–decay was tested by subsampling modern sites near LGM sites, showing near-identical decay patterns to the full dataset. Chronological robustness was assessed by stratifying LGM samples by age-control levels; anomalous high-similarity/large-temperature-difference pairs persist except in the lowest-confidence subset due to spatial distribution. Model simulations: Evaluated PMIP3/PMIP4 LGM equilibrium simulations (no freshwater hosing; fixed ice sheets) and their PI controls: NCAR-CCSM4 (two members), CNRM-CM5, FGOALS-g2, GISS-E2-R (two members), MIROC-ES2L, MPI-ESM1.2-LR, MRI-CGCM3, AWI-ESM. Additional AWI-ESM sensitivity experiments imposed freshwater fluxes (hosing) in the North Atlantic to reduce AMOC: (1) 0.2 Sv for 150 years offshore Newfoundland (50–70° N), (2) 0.2 Sv for 150 years across the subpolar North Atlantic (50–70° N), and (3) stochastic 0.1 ± 0.05 Sv for 1,050 years offshore Newfoundland without global salinity conservation. Temperatures were averaged over years 101–150 for short runs and the last 250 years for the long run. AMOC responses were quantified (default LGM ~18.2 Sv; hosing reduces to 11.8, 4.3, and 8.1 Sv, respectively). Analyses included: clustering assemblages (k-means) into cold/transitional/warm biogeographic groups; temporal turnover mapping; gridding the reconstruction (DIVA) for visualization; latitudinal patterns using generalized additive models with 95% CIs. All computations were done in R with established spatial and ecology packages.

- Modern core-top assemblages show strong similarity decay with increasing thermal distance; this pattern should hold for LGM if niches are stable, which is supported by 98% good modern analogues for LGM sites. - When LGM assemblages are paired with simulated LGM temperatures, many site pairs exhibit high assemblage similarity (>0.75) despite large simulated temperature differences (~10 °C), falling outside modern limits. This inconsistency appears across all assessed LGM simulations and is not due to model bias in PI controls or to chronological/sampling artifacts. - Biogeographic maps reveal cold-water assemblages expanded equatorward during the LGM, especially in the North Atlantic, at the expense of transitional assemblages; tropical assemblage extents changed only marginally, indicating a steeper meridional ecological gradient, particularly in the North Atlantic. - Foraminifera-based reconstructions (annual mean, 50 m) indicate global mean LGM cooling of −2.45 °C (95% CI: −2.57 to −2.33 °C) relative to modern, consistent in magnitude with model ensemble mean (−2.15 °C; range −1.30 to −2.89 °C). - Spatial patterns diverge markedly: reconstructed subpolar North Atlantic (40°–60° N) cooling averages ~−7.3 °C, whereas models simulate much more uniform cooling and substantially underestimate North Atlantic cooling. Model–data differences average up to 4.9 °C across 40°–60° N (ensemble mean) and up to 8.3 °C for individual models. - The pattern of large temporal turnover and strong cooling in the North Atlantic resembles an AMOC fingerprint; simulations appear too warm there and too spatially uniform globally. - Sensitivity simulations with reduced AMOC (via freshwater hosing) produce a colder subpolar North Atlantic, reduce the number of anomalous high-similarity/large-temperature-difference pairs in similarity–decay plots, and cut model–data temperature differences in the mid-latitude North Atlantic by up to ~50%. - AMOC magnitudes: default LGM ~18.2 Sv; hosing experiments reduce AMOC to 11.8 Sv (coastal), 4.3 Sv (subpolar), and 8.1 Sv (coastal long), aligning better with independent circulation reconstructions. - Robustness checks show that results are insensitive to analogue filtering, spatial sampling, and age-control uncertainties; reconstructed subtropical warming signals at some low-latitude sites are not artifacts of poor analogues.

The findings support an ecologically grounded diagnosis that current LGM simulations cool the oceans too uniformly and underestimate the steepening of meridional thermal gradients, particularly in the North Atlantic. The mismatch between simulated temperatures and assemblage-derived similarity–temperature relationships indicates that the spatial imprint of glacial cooling, rather than its global mean magnitude, is misrepresented in many models. The foraminifera biogeography and reconstructions point to a markedly colder subpolar North Atlantic consistent with a reduced or more southward-limited AMOC during the LGM. Sensitivity experiments confirm that weakening AMOC is a physically plausible mechanism to reconcile models with data, substantially improving both the similarity–decay consistency and the regional temperature fields. Because species’ thermal niches remained stable and nearly all LGM assemblages have modern analogues, the macroecological approach provides a robust, independent evaluation that sidesteps proxy seasonal/depth ambiguities and leverages spatial gradients directly. The work underscores that accurate simulation of spatial heterogeneity and oceanic thermal gradients is essential for credible past climate modelling and, by extension, for confidence in future projections where regional patterns matter for ecosystems and societies.

The study introduces and applies a macroecological, similarity–decay framework to evaluate LGM climate simulations using planktonic foraminifera assemblages. It demonstrates that while global mean cooling is broadly captured, models generally fail to reproduce the steep meridional gradients inferred for the North Atlantic, indicating much stronger regional cooling during the LGM likely associated with a reduced AMOC. Freshwater-hosing sensitivity experiments support this mechanism and improve model–data agreement. The approach offers a robust way to diagnose past climate simulations without relying solely on temperature proxy calibrations and can be extended to intervals with extinct taxa. Future work should focus on resolving oceanic thermal gradients in both equilibrium and transient LGM simulations, better representing AMOC dynamics, and expanding macroecological evaluations across additional proxy communities and regions.

- Although nearly all LGM assemblages have good modern analogues and niches appear stable, the approach assumes stability of species’ thermal niches and that temperature remains the dominant driver; potential collinearity with other environmental variables cannot be completely ruled out (though a systematic global change is considered unlikely). - Transfer-function reconstructions, while robust to analogue filtering and with spatially autocorrelated uncertainties accounted for (median 2σ ≈ 1.52 °C), still carry regional uncertainties and depend on the chosen calibration (annual mean, 50 m). - Chronological uncertainties exist in parts of the LGM dataset, though sensitivity analyses indicate the key patterns are robust and dominated by high-confidence samples. - Model-side uncertainties include structural and setup differences (e.g., ice-sheet configurations, mixing), and freshwater-hosing experiments, while illustrative, are idealized and not unique in mechanism (other processes like ice-sheet height or mixing changes could also yield cooling). - Spatial coverage remains limited to available sites; comparisons are constrained to those locations, and gridded fields for visualization may smooth local variability.