Climate Model Synthesis, or CMS, is Berkeley Earth’s framework for transforming climate model projections into local, observation-constrained temperature information.
The system combines more than 400 climate model runs from 45 different models, including historical simulations and projections across five SSP scenarios. These model simulations are compared with Berkeley Earth observations, the Berkeley Earth High Resolution Data Set, ERA5, and Berkeley Earth’s weather station records. The resulting products provide bias-corrected and statistically downscaled estimates of historical and future temperature change through 2100.
All gridded model data are prepared at 0.25° × 0.25° latitude-longitude resolution.
Overview
CMS is designed to preserve the strengths of global climate models while making their results more useful for local analysis. Climate models provide physically based projections of long-term climate change, but their raw output may contain regional biases, coarse spatial resolution, and imperfect representations of local variability and extremes.
CMS addresses these challenges through a sequence of model decomposition, observational comparison, bias correction, statistical downscaling, model weighting, and weather simulation. The result is a bias-corrected and downscaled system that provides information on past and future changes in local weather, consistent with historical observations and with the uncertainty represented in the CMIP6 ensemble.
Decomposing Temperature Behavior
For each model, temperature time series are separated into three components:
- Long-term trend
- Seasonality
- Short-term weather variability
This decomposition allows each part of the temperature record to be analyzed and corrected in a way that is appropriate to its behavior.
The long-term trend is modeled using a 30-year LOESS fit. This captures slow changes in the climate system while smoothing over short-term weather variability.
After removing the long-term trend, seasonality is modeled using a 30-year windowed sinusoidal regression. This regression represents the annual cycle and higher-order seasonal modes, allowing the seasonal structure of temperature to vary over time.
The remaining residuals are treated as short-term weather variability. Weather is represented as an autocorrelated probability distribution with generalized Pareto tails. This allows CMS to represent ordinary day-to-day variability as well as extremes. The weather distribution is allowed to vary both seasonally and over the long term.
Observational Comparison and Bias Correction
Each component of each model is compared to corresponding decompositions of observational and reanalysis-based data, including:
- Berkeley Earth High Resolution Data Set
- ERA5
- Berkeley Earth weather station data
Biases in historical mean temperature are corrected using observational data. In addition, biases in long-term trends are adjusted using a hierarchical rescaling procedure in regions where a model disagrees with the historical data at greater than 95% confidence.
Weather distributions are corrected using quantile mapping. This allows the model’s residual weather variability, including the shape of the distribution and the behavior of extremes, to be brought into better agreement with observational data.
These adjustment procedures gently nudge each model toward better agreement with observations during the historical period. They are designed to reduce systematic errors while preserving the climate change information contained in the model simulations.
Statistical Downscaling
Climate models differ in their native spatial resolution, and many are too coarse to directly represent local climate conditions. CMS therefore statistically downscales model output to a common 0.25° × 0.25° latitude-longitude grid.
The downscaling uses Berkeley Earth and ERA5 data as guidance, generating 0.25° component fields for trend, seasonality, and weather variability. This creates a consistent high-resolution representation of each model while maintaining the large-scale climate behavior projected by the model.
Model Reliability and Weighting
Even after mean and long-term trend adjustments, important differences remain between models and observations. These remaining differences are used to assess the relative reliability of each model in each region.
Within each scenario pool and each region, approximately half of the models are selected for downweighting because of relatively poor performance over the historical period. This does not eliminate the diversity of the model ensemble, but it reduces the influence of models that are less consistent with observed regional climate behavior.
The weighting procedure helps CMS avoid a simple equal-weighted average of all available models. Instead, the synthesis gives greater influence to models that better reproduce historical conditions, while still preserving the spread of plausible futures represented across the CMIP6 ensemble.
Building the Superensemble
For each scenario, the adjusted and downscaled models are combined with appropriate weights to estimate both the mean behavior and the variation across the model superensemble.
CMS uses Cholesky decomposition to capture variations across models in the trend, seasonality, and weather domains. This preserves important correlation relationships among the components of temperature behavior and across the ensemble.
The resulting representations are combined into a full model of long-term climate change that spans the behavior of the bias-corrected CMIP6 models. This representation includes both the central tendency of the ensemble and the uncertainty implied by differences among models.
Weather Simulation
The combined CMS representation can be treated as a weather emulator. It allows Berkeley Earth to generate large numbers of samples from the associated probability distributions, producing simulated daily temperature time series for any location on Earth.
These simulated time series include TAVG, TMAX, and TMIN. They are designed to preserve correlations, variability, seasonality, long-term change, and extreme behavior.
Because CMS can generate thousands of plausible daily temperature time series, it can estimate both average changes and probabilistic changes in local extremes. This provides a practical way to translate climate projections into quantities that are useful for planning, risk assessment, and adaptation.
Derived Metrics
Many useful climate statistics are not directly available from annual or monthly average temperature alone. They must be calculated from daily weather behavior.
CMS uses simulated daily temperature time series to derive metrics such as:
- Cooling degree days
- Heating degree days
- Days above selected thresholds, such as 35 °C
- Annual maximum temperature
- Changes in the probability of extreme heat events
These derived metrics provide a more detailed view of local climate change than mean temperature alone. They can help describe changes in energy demand, heat stress, infrastructure risk, and other temperature-sensitive outcomes.
Berkeley Earth looks forward to working with partners who can make use of simulated daily time series to prepare custom statistics suited to other applications. Many users may need specialized thresholds, sector-specific indices, or metrics tailored to particular decisions. CMS is designed to support that kind of custom analysis.
Summary
Climate Model Synthesis provides a bias-corrected, statistically downscaled, observation-constrained representation of past and future local temperature change.
By decomposing temperature into trend, seasonality, and weather variability; comparing each component to observations; correcting systematic biases; downscaling to 0.25° resolution; weighting models by regional performance; and generating simulated daily time series, CMS creates a flexible framework for understanding local temperature change through 2100.
The result is a system that is consistent with historical observations, informed by the CMIP6 model ensemble, and capable of describing both average climate change and changes in local extremes.
