Berkeley Earth Climate Model Synthesis

The global cimate models (GCMs) included in the CMIP6 compilation provide critical insights into global warming but often exhibit biases at local and regional scales. While bias-correction and downscaling are common, existing methods rarely incorporate multi-model ensembles or extend their applicability to a global field. We present a downscaled and bias-corrected CMIP6 synthesis that spans the full range of CMIP6 projections to enhance the accuracy of future climate projections, facilitate adaptation efforts, and increase climate resilience. 

Our new work synthesizes a bias-corrected and downscaled surface temperature product derived from 45 GCMs and 374 runs across five shared socioeconomic pathways. We employ a robust bias-correction framework that compares historical model runs against reanalysis data (ERA5: 1940-present) and observation-based data (Berkeley Earth: 1850-present). Each grid cell is decomposed into three components: (1) long-term trends, (2) annual seasonality, and (3) short-term weather variability.  Trends and seasonality are calculated with a LOESS fit with Gaussian weighting.  Residual daily variability is represented using evolving probability distributions with explicitly modeled extremes through generalized Pareto formalism to enable accurate estimation of rare events.  The resulting fields are then statistically downscaled to 0.25° x 0.25° latitude-logitude resolution using predictive regressions derived from high resolution historical observations.  Further, each component is bias corrected and scaled to match the mean and trends observed in the historical period.

This analysis results in bias corrected and downscaled verions of each of the input GCMs.  These bias corrections, based on constraints over the historical period, significantly reduce the spread in model projections of the future, and by extension the implied uncertainty in long-term warming scenarios.

The corrected models are then further synthesized into a unified dataset, with model selection and weighting guided by historical accuracy.  This provides easy access to local changes, consistently representing both past temperature changes and expected future changes.  Data is provided both for long-term trends as well as daily extremes with measures of uncertainty guided by the remaining variations across models.  Further, this approach makes it easy to calculate changes in temperature derived variables, such as cooling-degree days or heat wave indices.

The Berkeley Earth climate model synthesis will deliver detailed probabilities of extreme climate events for each shared socioeconomic pathway.  The initial focus is on surface temperature changes, with an additional synthesis of precipitation changes planned for the next phase of this work.  This product aims to support future academic research, inform policy and adaptation strategies, and provide actionable climate risk insights for asset management and decision-making.