Certainty and uncertainty

What are the sources of uncertainty in climate projections?

Uncertainty about future climate projections comes from several sources. Here these sources of uncertainty are broken down into three main categories, a different but similar way to categorise these sources of uncertainty can be found in Foley (2010). The first source, and one of the largest unknowns, is the future emissions of greenhouse gases and aerosols. Since this uncertainty is impossible to quantify mathematically, it is presented as a series of possible emission scenarios or projections. These scenarios are then used in GCM simulations to study the impact on climate.

The second and the third sources of uncertainty deal with the response of the physical system to the increase in greenhouse gases and aerosols. Specifically, the second source is a large scale response to changes in atmospheric constituents. It can be sampled by using different GCM (“model structural uncertainty”), and different parametrisations within a single GCM (“model parametric uncertainty”). The third source is a local response given a large-scale response. In the case of RCMs this includes the uncertainty in model physics and structure similar to issues associated with GCMs, while for statistical downscaling this includes uncertainties associated with the statistical technique used. In combination these sources of uncertainty provide a limit to the confidence that can be placed in any particular projection of future regional climate.

How is uncertainty address in regional climate modelling?

Quantifying this uncertainty is done by creating a collection, or ensemble, of climate simulations that sample various parts of the uncertainty described above. Emission scenario uncertainty is addressed by running simulations from more than one scenario. To quantify the structural uncertainty associated with GCMs an ensemble of many GCMs should be used and similarly for RCMs (or dynamical downscaling) many RCMs should also be used. Ideally these GCMs and RCMs would be independent of each other ensuring they are sampling different parts of the plausible future climate space. Once an ensemble sampling these uncertainties has been established there are multiple methods for combining the information to establish a probabilistic future climate change prediction. Déqué and Somot (2010) used a technique that weighs a frequency distribution based on model performance. Bayesian analysis has also been used in a number of ways (Tebaldi et al., 2004,2005; Buser et al., 2010) and is an area of active research.

How was uncertainty addressed in the NARCliM modelling?

Due to the many uncertainties involved in producing climate projections, NARCliM is providing an ensemble of 12 simulations (rather than a single projection).

A dynamical regional climate model (WRF) was used to dynamical downscale GCM projections. Three configurations of WRF were run with four separate GCMs to produce an ensemble of 12 member runs, each using the IPCC SRES A2 emission scenario. This emission scenario was used because it reflects recent trends in global emissions1, 2. An ensemble approach was used to provide robust regional climate projections which span the range of likely future changes in climate for south-eastern Australia.

To minimise potential bias from a single model, four GCMs from the World Climate Research Program (WCRP's) Coupled Model Inter-comparison Project phase 3 (CMIP3) suite of models were used. CMIP3 modeling informed the IPCC Third Assessment Report (TAR) and Fourth Assessment Report (AR4), and has been used together with CMIP5 modeling to inform the IPCC’s Fifth Assessment Report (AR5). The four GCMs (MIROC3.2, ECHAM5, CCCMA3.1 and CSIRO-MK3.0) were selected based on their performance compared to observations for the recent past, independence and ability to span the range of plausible future climates3.

WRF was run with different settings which reflect uncertainties in our understanding of some physical processes. Varying these settings between model runs and using several different settings with each GCM provides a more realistic set of projections. Performing multiple model runs also captures more reliable information on rare, extreme weather events, such as heatwaves, heavy rain and drought. Initially the WRF model was run with 36 different physic scheme setting, with model performance tested for five climate variables. On the basis of model performance and independence, three WRF model configurations were selected for use in the downscaling4.

Interpreting Future Climate Projections

Using ensemble projections

Due to the many uncertainties involved in producing climate projections, NARCliM is providing an ensemble of 12 simulations (rather than a single projection). Each of these simulations represents an “equally plausible” future climate, with the range of future changes within the ensemble providing a measure of the confidence in any particular change.

The use of all 12 ensemble members is recommended to estimate climate change impacts, so ensuring an “equally plausible” range of possible impacts. If this is not possible (due to logistical constraints etc.) then you could choose a smaller number of simulations (2 or 4) that span the range of the climate variable for which your system is most sensitive.

You could also use the central estimate (multi-model mean) provided keeping in mind the confidence range of the full ensemble.

What is the central estimate and how should it be used?

Central estimates of future climate change are used to summarise changes projected by the 12 modeling simulations undertaken for NARCliM. Multi-model median and multi-model mean are commonly used internationally to provide a reasonable “middle” projection.

The multi-model mean method was used to provide a central estimate for the NARCliM projections because of the model independent measure used to select the four GCMs. GCMs that are most independent contribute the most independent information in calculating a multi-model mean. This means that the multi-model mean of a selection of highly independent GCMs from a larger ensemble of GCMs (e.g. the CMIP3 ensemble) should be closer to the multi-model mean of the larger ensemble than the multi-model mean of a selection of GCMs that are less independent.

When using central estimates it is important to note the level of model agreement on the direction of change. If models are projecting both increases and decreases, the multi-model mean is likely to be within the range of model uncertainty and needs to be interpreted with caution. In such cases, reference should be made to maps showing the suite of individual simulation projection results and to location-specific bar graphs showing the span of plausible changes.

What does model agreement mean?

Model agreement refers to the number of individual model simulations within the ensemble that agree on the direction of change (e.g. increase or decrease in rainfall), and is used to determine the level of confidence in the change projected. The more models agree, the greater the confidence in that direction of change.

For example, if all simulations project a rainfall decrease for an area then this decrease has high confidence. If 9 of 12 simulations project a rainfall decrease for an area then this decrease has medium confidence.


  1. Peters GP, Andrew RM, Boden T, Canadell JG, Ciais P, Le Quere C, Marland G, Raupach MR, Wilson C. Commentary: The challenge to keep global warming below 2 degrees C Nature Climate Change 2013, 3:4-6.
  2. Rahmstorf S, Cazenave A, Church JA, Hansen JE, Keeling RF, Parker DE, Somerville RCJ. Recent climate observations compared to projections Science 2007, 316:709-709.
  3. Model independence means that the chosen models do not exhibit the same strengths and weaknesses in their representation of the climate.
  4. Evans, J.P. and Ji, F. 2012, Choosing GCMs, NARCliM Technical Note 1, 7pp, NARCliM Consortium, Sydney, Australia