Step 6: Select Model Simulations
Not all available model data will best meet your needs. Based on your project requirements (step 1) and climate model data evaluation (step 5), you should now select the climate model simulations you will use for subsequent analyses.
For example, simulation data from more than 40 global climate models (GCMs) is available from the CMIP5 archive. However, due to resource constraints, you may have decided in step 1 to analyze simulation data of just three to five GCMs. In this step, you can select those model simulations based on a literature review about CMIP5 GCM performance in your region, if available, or on the results of your evaluation (step 5). Alternatively, you may select three GCMs that represent a range of plausible future projections (e.g. dry case, medium case, and wet case) to conduct subsequent steps. (See Climate Futures for more detail.)
The same approach can also be used to select GCM data that will be downscaled if you need to conduct and analyze high-resolution regional climate model (RCM) simulations. High-resolution RCM simulations are computationally intensive (e.g. one dynamically downscaled 100-year simulation may take three or four months to complete on a supercomputer) so a compromise is necessary.
In some cases, even if all modelled future climates could be produced, it is likely that the end user would have limited capability to run the impact model for all cases and then interpret the results. Therefore, it is often necessary to select a sub-set of available models to provide regional projections, using a selection process that adequately captures the full range of possible futures.
When selecting model data for developing climate projections or scenarios there are a number of important considerations.
Choose all necessary variables from one model
Do not combine variables from different models. For example, do not use a combination of rainfall projections from model A and temperature projections from model B. All projections of rainfall and temperature change applied in an impact model should have a consistent set of assumptions, including the choice of climate models, time period, and greenhouse gas emissions scenario. Arbitrary mixing-and-matching of projections degrades the realism of the outcome and limits comparability of different impact and risk assessments. Variables such as temperature, rainfall, evaporation, and humidity are highly correlated, meaning a change in one variable influences other variables.
In addition, estimates of impacts should be calculated independently for each of the selected climate models, and should not be based on the average of the selected climate models. This is to provide estimates that are representative of plausible climate futures which can then be ranked according to their relative impact.
Select a range of models that capture the spread of all plausible scenarios
At a minimum, select a worst case, best case, and the case with the most consensus (based on all available model projections)—even if this means selecting a lower resolution projection (such as a GCM). An example of a tool to do such a selection is the Climate Futures tool.
Try to select an ensemble (group of models) suitable for the purpose
Choose resolution, time period, and ensemble size to match with available datasets and computer resources, so your application is achievable.
Include information likelihood (through the percent of models indicating a particular direction and/or magnitude of projection)
The percent of models giving certain projection is one measure of likelihood or risk.
Consider model evaluation results
- Ensure that the models capture key atmospheric features, seasonality of temperature and rainfall, and variability over time, such as related to El Niño–Southern Oscillation.
- Individual models perform better for some variables than other models and rarely does one single model consistently outperform all other models for all variables. However, if a model does not capture some key atmospheric feature of the current climate, it is unlikely to capture that feature in the projections, so you may wish to disregard that model.
- You may also choose to exclude models that perform badly across a range of metrics, although it is generally best practice to use as many models as possible.