Paleoclimate data provide constraints on climate models’ large-scale response to past CO₂ changes

Climate models are frequently used to project future climates under CO₂ concentrations exceeding those of the recent observational period. However, traditional model evaluation often overlooks the potential of paleoclimate proxy data. Paleoclimate simulations offer a unique opportunity to test model behavior across a wide range of forcings, encompassing those expected in the coming centuries. A model accurately simulating both past and present climates inspires greater confidence in its future projections. This study focuses on three periods characterized by substantial CO₂ forcing: the LGM (~180 ppmv CO₂), the MPWP (~400 ppmv CO₂), and the EECO (~1500 ppmv CO₂). These periods have been extensively studied within the Paleoclimate Modelling Intercomparison Project (PMIP), providing a rich dataset of model simulations. The selection also allows for direct relevance to future projections due to the substantial CO₂ changes involved. The study examines three large-scale properties: global mean surface temperature (GMST), polar amplification, and land-sea warming contrast. GMST is a fundamental metric, key to international climate agreements. Polar amplification, the disproportionate warming in polar regions, is critical given its impact on sea-level rise. Land-sea warming contrast, the difference in warming between land and ocean, is linked to changes in the hydrological cycle and atmospheric circulation. While defining these metrics in models is straightforward, their estimation from paleoclimate proxy data is challenging due to data sparsity and uncertainties.

Previous research utilizing paleoclimate data in climate model evaluation has yielded mixed results. Some studies have found a clear correlation between modern climate sensitivity (ECS) and paleo GMST, while others have not. The uncertainty in CO2 concentration estimates for the Pliocene and EECO, compared to the LGM, adds complexity to evaluating climate sensitivity. Prior work on polar amplification has focused on different proxies and metrics, often using deep-ocean temperatures rather than surface temperatures. Studies on land-sea warming contrast have highlighted the challenges in reconciling model simulations with observational data, especially given the complex interplay between atmospheric circulation, hydrological cycle and surface properties. The current study builds upon this existing research by employing a standardized set of metrics and a broader range of models and proxy data to provide a more comprehensive evaluation of climate models' performance.

The study evaluates climate models using metrics derived from paleoclimate proxy observations. Three large-scale properties are considered: GMST, polar amplification, and land-sea warming contrast. Two types of metrics are defined: "true" metrics based on globally defined fields and "site-specific" metrics based on the locations of proxy data. For GMST, the study utilizes IPCC AR6 assessed values for the three paleo periods and also includes the GMST metric of Annan et al. (2022). For polar amplification and land-sea warming contrast, site-based proxy data is employed, drawing from established syntheses and compilations for each time period. Data sources include those for LGM sea surface temperatures (SSTs) and land air temperatures (LATs); MPWP SSTs and LATs; and EECO SSTs and LATs. Specific references for each dataset are provided in the original paper. Model simulations come from the fourth and third phases of the Paleoclimate Modelling Intercomparison Project (PMIP4 and PMIP3). The standard boundary conditions (CO2, non-CO2 greenhouse gases, ice sheets, vegetation) and simulation protocols (run length, initial conditions) are detailed in the original paper. Both PMIP4/CMIP6 and PMIP3/CMIP5 simulations are included. The paper meticulously describes the model selection criteria, including those excluded due to different reference frames or CO2 levels in the EECO simulations. The analysis compares model-simulated metrics with proxy-based metrics, accounting for uncertainties in both. The study explicitly defines the methods for calculating each metric (true and site-specific), addressing the challenges posed by sparse and uncertain proxy data. Statistical comparisons are used to assess the degree of agreement between model simulations and observations.

The study found that the CMIP6/PMIP4 multi-model mean GMST shows remarkably good agreement with proxy data for all three paleo periods, representing a temperature range of about 20°C. This agreement is an improvement over CMIP5/PMIP3. The LGM exhibits a high signal-to-noise ratio, making it a stringent test for model-data comparison. However, significant inter-model spread exists, with many individual models falling outside the observed GMST range. There is a general indication that models with ECS values outside the IPCC-assessed range (2-5°C) simulate paleo temperatures inconsistent with proxy data. High-ECS models simulate excessive warming, while low-ECS models show insufficient warming. The results suggest that paleoclimate simulations might be better discriminators of high- and low-ECS models than historical simulations, potentially because paleoclimate simulations are closer to equilibrium. However, further research is needed to confirm this finding. The study also examined polar amplification. While models qualitatively reproduce polar amplification, quantitative discrepancies emerge, especially in the EECO. The EECO proxy-based polar amplification is much higher than that simulated by the models, possibly due to warm biases in proxy SST temperatures from the southwest Pacific. When these data are excluded, model-data agreement improves. Land-sea warming contrast is reasonably well-simulated in the LGM, but discrepancies are seen in the MPWP and EECO. Models consistently show a saturating relationship between GMST and land-sea warming contrast, consistent with theoretical expectations. The discrepancies are, in part, associated with regional biases in the proxy data, particularly in the northern hemisphere continental interiors and the southwest Pacific.

This study's findings highlight the value of paleoclimate data in constraining climate models. The improved agreement between CMIP6/PMIP4 models and proxy data suggests progress in model development. The LGM, with its high signal-to-noise ratio, proves a powerful benchmark for model evaluation. The discrepancies between model simulations and proxy data for polar amplification and land-sea warming contrast point to areas needing further improvement in models, particularly concerning the representation of cloud feedbacks, ocean circulation, and regional climate processes. The study identifies specific regions where proxy data uncertainties warrant further investigation, notably the MPWP high latitude continents and the EECO southwest Pacific. These uncertainties affect the robustness of the proxy-based metrics.

This research provides a framework for evaluating climate models using paleoclimate data, particularly emphasizing the importance of incorporating multiple lines of evidence. The CMIP6/PMIP4 models show improved performance relative to their predecessors. The study reinforces the need for continued refinement of climate models, especially in areas where model-data discrepancies exist, and emphasizes the vital role of paleoclimate data in model development and validation. Future research should focus on resolving uncertainties in proxy data and improving model representation of key physical processes.

The study acknowledges limitations inherent to paleoclimate data, including sparsity, uncertainties in proxy reconstructions, and potential biases. Regional differences in data density and uncertainties, such as those seen in the EECO southwest Pacific and MPWP high-latitude continental interiors, introduce variability in the proxy-based metrics. The differing temporal windows of terrestrial and marine proxies in the EECO also introduce uncertainties. Furthermore, the study notes that some proxy sites are located at high elevations not fully resolved by the models, and in the EECO, paleoelevation is uncertain for some locations. The wide temporal window of the EECO (approximately 4.1 Myr) means that orbital forcing and CO2 variations might influence the proxy signal.