During the Last Deglaciation, global surface temperature rose by about 4–7 °C over several millennia. We show that changes in year-to-year up to century-to-century fluctuations of temperature and precipitation during the Deglaciation were mostly larger than during either the preceding or succeeding more stable periods in 15 climate model simulations. The analysis demonstrates how ice sheets, meltwater, and volcanism influence simulated variability to inform future simulation protocols.

Earth's past temperature reconstructions are critical for understanding climate change. We test the ability of these reconstructions using climate simulations. Uncertainties, mainly from past temperature measurement methods and age determination, impact reconstructions over time. While more data enhance accuracy for long-term trends, high-quality data are more important for short-term precision. Our study lays the groundwork for better reconstructions and suggests avenues for improvement.

The ability of climate models to faithfully reproduce past warming episodes is a valuable test considering potentially large future warming. We develop a new method to compare simulations of the last deglaciation with temperature reconstructions. We find that reconstructions differ more between regions than simulations, potentially due to deficiencies in the simulation design, models, or reconstructions. Our work is a promising step towards benchmarking simulations of past climate transitions.

Data–data and data–model vegetation comparisons are commonly based on comparing single vegetation estimates. While this approach generates good results on average, reducing pollen assemblages to single single plant functional type (PFT) or biome estimates can oversimplify the vegetation signal. We propose using a multivariate metric, the Earth mover's distance (EMD), to include more details about the vegetation structure when performing such comparisons.

ll-posedness of the inverse problem and sparse noisy data are two major challenges in the modeling of high-dimensional spatiotemporal processes. We present a Bayesian inference method with a strongly regularized posterior to overcome these challenges, enabling joint state-parameter estimation and quantifying uncertainty in the estimation. We demonstrate the method on a physically motivated nonlinear stochastic partial differential equation arising from paleoclimate construction.