Over the past decade, the ocean carbon sink absorbed 29% of total anthropogenic CO₂ emissions. To quantify this sink from sparse observations, Global Ocean Biogeochemical Models (GOBMs) and Earth System Models (ESMs) simulate the global ocean carbon cycle, typically at 1-by-1-degree horizontal resolution. Recent assessments of observation-based ocean carbon data products use GOBMs and ESMs as “testbed” truths: data products sample the testbeds at the spatio-temporal coverage of independent observations, then reconstruct surface ocean pCO₂ using statistical and machine learning techniques. The testbed approach then allows for a comprehensive assessment of the data product’s ability to reconstruct the full field in the context of sparse pCO₂ data coverage. It has not been studied whether differences in the testbed model skill in representing real-world mean and seasonality of pCO₂ may impact testbed results. Thus, this study assesses the data products by ensemble-averaging based on their testbed model choice. We use 5 GOBMs over 2001–2021 and 18 ESM members over 1982–2023. Considering regional, basin, and global scales, the bias, unbiased RMSE, and correlation are calculated of each data product against its testbed. Likewise, the same metrics are computed for each testbed model against real-world independent data from GLODAP, SOCAT, and LDEO pCO₂ databases. Results show that sorting the testbed models based on their skill against observations has a minor impact on the assessment of reconstruction skill across 7 data products. Echoing published findings, ensemble-averaging generally improves global and regional reconstruction skill compared to using single-member testbeds. Nonetheless, the equatorial and Southern Ocean subtropical biomes show greater sensitivity to testbed choice, where including certain testbeds can marginally worsen, but not improve, ensemble-averaged skill. This study highlights the versatility of testbeds and endorses the use of ensemble-averaging in multi-model testbed experiments, since specific testbed model choices can harm regional reconstruction skills, especially in observation-scarce and biogeochemically complex biomes.

AGU 2026 Annual Meeting, San Francisco · Session A072 — From Physics to AI: Benchmarking, Bias Diagnosis, and Process Representation in Earth System Models.
Anna Chen, Galen A. McKinley, Thea Hatlen Heimdal, Amanda R. Fay.

The Last Interglacial (LIG, 130-115 ka) was the most recent interval during which global temperatures were warmer than today. LIG sea surface temperature (SST) variability offers a natural analog for evaluating ocean-climate responses under ongoing warming. Utilizing the best-current proxy-based dataset of 104 temporally interpolated timeseries over 130-115 ka from Hoffman et al. (2017), we provide a preliminary global field reconstruction of LIG SST at 0.1 kyr intervals and 1-by-1-degree horizontal resolution using Reduced-Space Optimal Interpolation (RSOI). We perform Empirical Orthogonal Function (EOF) analysis on annual modern SST fields from the HadISST v1.1 dataset (1870-2020). Treating the Hoffman et al. dataset as independent observations, weight vectors are computed by RSOI from a 20-dimensional “reduced” space, so that the full field SST is reconstructed as a weighted sum of the 20 leading EOFs. The estimated grid wise uncertainties and uncertainty covariances are calculated from the proxy uncertainties provided by Hoffman et al. and the truncation error of full dimensional SST to the reduced space. By iterating the RSOI procedure, the mean and covariance of the SST field converged by the 5th iteration. In the resulting reconstruction, the first 20 eigenvalues explain 88% of the total variance in centennial means. The spatio-temporal evolution of the reconstructed SST generally matches published literature. The Northern Hemisphere extratropics show the greatest sensitivity to both warming and cooling. The tropics have the lowest SST anomaly throughout the LIG. Maximal global mean SST of +0.2 deg C from the 1870-2020 mean is reached around 126 ka, and near-modern global mean SST is reached by 121 ka. Overall, the RSOI method provided a computationally efficient, well-constrained, and high-resolution solution, complete with its uncertainty estimates. For the planned further assessment of reconstruction’s fidelity, we mined the contributor data files from the NOAA Paleoclimatology database, resulting in finding 62 records of proxy-based LIG SST estimates that are independent of the Hoffman et al. compilation.

AGU 2026 Annual Meeting, San Francisco · Session PP008 — Late Quaternary Climate and Environmental Change: Proxies, Processes, and Models.
Anna Chen, Alexey Kaplan, Jerry F. McManus.

Carbon fixation by marine phytoplankton is crucial in influencing the pelagic and global carbon cycle. Future climatic changes likely affect phytoplankton growth and the net primary production (NPP) of surface ocean. However, the magnitude and directionality of growth changes are poorly constrained. Since phytoplankton are highly adaptable, understanding future growth in lab-based experimentation requires accurate current and future conditions. Hence, this study seeks to constrain the temporal and spatial variability of two important physical factors (temperature, CO₂) in areas of high NPP. Historical ocean conditions were examined by combining observational data and modeled products. For growth estimates, the author utilized satellite data from the Carbon-based Production Model, a monthly and global estimate of marine NPP at 1/6º-by-1/6º grids. Surface ocean temperature was provided by the World Ocean Atlas, and the partial pressure of CO₂ (pCO₂) from the Global Ocean Surface Water pCO₂ Database. The author calculated regional means and weighted conditions by NPP to evaluate how marine conditions with high NPP deviated from regional averages. Results across 14 regions show that, apart from the Southern Ocean, summer phytoplankton blooms mostly occur in areas with lower temperatures and lower pCO₂ conditions relative to regional averages. Opposingly, winter phytoplankton growth occurs in areas with higher temperatures and pCO₂. These findings provide insights for lab-based experiments. Rather than viewing growth uniformly across time and space, researchers should focus on the deviation of in-situ conditions in high-growth areas from their seasonal and regional averages to reflect temporally- and spatially-varied growth, outlining more accurate climatological contexts for future lab-based experimentation.

Mentors: Dr. Sarah Hurley, Dr. Henry Holm. Laidlaw Undergraduate Research and Leadership Scholarship, Summer 2024.