2.3 Climate Model Uncertainty#
From the discussion in the previous section, you likely picked up on the fact that projections made using climate models are inherently subject to uncertainty. In fact, each of the three main sources of uncertainty in climate modeling were strongly hinted at in Sections 2.1 and 2.2, but here they will be laid out explicitly, and we will discuss how climate scientists attempt to quantify the effects of each source of uncertainty.
2.3.1 Scenario Uncertainty#
The most obvious source of uncertainty in future climate projections is that we don’t know what actions humans will take that affect the climate. We don’t know how much CO\(_{2}\) and other greenhouse gases we will emit; we don’t know how land use will change, and we don’t know what sorts of policies will be adopted and implemented to mitigate climate change. In other words, there is no “correct” set of boundary conditions to give to a climate model to do projections of future climate. Instead, we can study a range of plausible future pathways of emissions and other anthropogenic factors that influence the climate, and take a conditional approach. If humans follow this certain course of action, then this is what the effects on the climate could be. As a part of the CMIP protocols and the IPCC Assessment Reports, experts develop a set of scenarios reflecting different courses of action, from aggressive mitigation and carbon removal to laissez-faire regulation and continued heavy reliance on fossil fuels. These scenarios include time series of greenhouse gas concentrations, land usages, and other human-influenced boundary conditions to use for future climate simulation.
|
|---|
Atmospheric CO2 concentrations for each SSP scenario, from Wikipedia [Shared Socioeconomic Pathways, 2023]. |
For CMIP5/AR5 (circa 2013-2014), these scenarios were called the Representative Concentration Pathways, or “RCPs”, each numbered by the year-2100 net top of atmosphere radiative forcing in units of \(W/m^{2}\) (RCP2.6, RCP4.5, RCP6.0, and RCP8.5). For the next phases of each of these projects (CMIP6 and AR6), these scenarios were expanded upon with the Shared Socioeconomic Pathways (SSPs), which include additional GHG scenarios as well as new factors such as population growth and economic factors so that social scientists can use the same analysis framework as physical climate scientists.
The existence of these different scenarios is great, but it doesn’t resolve the issue that we don’t (and can’t) know which one will be closest to reality. No likelihood of any of the SSPs or RCPs is quantified. That leaves you as someone studying the impacts of climate change with a choice. You can either look at an envelope of possible outcomes by choosing a high and low forcing scenario, be conservative and focus only on the worst-case, highest forcing scenario, or you can sidestep the issue entirely, and instead look at impacts at a particular level of global warming (i.e. 2\(^{\circ}\)C), using data from several different scenario-simulations together. The idea behind this latter approach is that the impacts of a certain amount of global warming should be similar regardless of how long it took to get there, and the main difference between scenarios is how long it takes to reach a certain warming level. Nothing guarantees this to be true, which is why it’s important to use data from multiple scenarios to construct the projections. Note that for some applications, such as the likelihood of record-breaking extreme events, the rate of warming can be relevant for the results [Fischer et al., 2021], so take this into account when deciding which approach to take regarding scenario uncertainty.
2.3.2 Internal Variability#
Section 2.1.2 introduced you to the concept of natural climate variability and broke it down into two components. Natural external forcings, such as variations in the intensity of incoming solar radiation, and internal variability, due to the inherent chaos in the climate system and modes of variability such as the El Nino-Southern Oscillation. For a given, fixed external forcing (be it natural or anthropogenic), there exist many possible time evolution trajectories for the system to take. Some possible realizations may have more El Ninos (for example), which would make them warmer than those that have more La Ninas. Looking at only one projection of the future, we can’t separate the effects of external anthropogenic forcing from the effects of internal variability. The amount of projected global warming (or any change to any quantity) could be inflated or deflated by the effects of internal variability. The only way to study the effects of internal variability is to run an ensemble of simulations, to sample the range of possible trajectories for a given pathway of external forcing. The spread across the ensemble (standard deviation, IQR, etc.) is how we quantify the uncertainty due to internal variability.
The particular trajectory for a given climate model simulation with fixed external forcing depends on the initial conditions. You may wonder why we can’t just initialize a model with observed initial conditions, to see what future the model projects for the trajectory of our real-world climate system. Doing this would essentially be an attempt to forecast the weather years or decades into the future - it’s impossible! We don’t have spatially complete and infinitely precise observations of the climate system, so after a very short period of time, the model trajectory would become completely detached from our real-world realization of internal variability. Instead, we estimate the size of the fluctuations due to internal variability by running an ensemble of many different simulations with the same external forcing, but slightly different initial conditions. Each individual simulation is called an ensemble member. Each ensemble member represents an equally plausible trajectory of the climate given the external forcing - even historical forcing. Climate modelers run ensembles for historical (20th century to early 21st century) and pre-industrial control levels of GHGs and other boundary conditions, to establish the effects of internal variability on the baseline climate from which anthropogenically forced changes are assessed. Ensembles of future forcing scenarios (the RCPs or SSPs) are used to disentangle true climate change signals from fluctuations due to internal variability. The best estimate of a model’s projected climate can be taken to be the ensemble mean, or average of results across all members of the ensemble.
Relevant for climate impact studies, internal variability causes more uncertainty for regional climate projections than for projections at larger scales. This is illustrated in the figure below, which shows the ensemble mean (line) and standard deviation (shading) of 2-metre air temperature projections for both Toronto (the single grid cell nearest to the city) and the global mean, from the CESM1 Large Ensemble project [Kay et al., 2015].
|
|---|
Ensemble mean and spread for annual mean 2 m air temperature projections, for a single grid cell over Toronto, and the global average. The future forcing scenario for this ensemble of simulations is RCP8.5. |
The shaded range, representing the spread across the ensemble members, is much larger for Toronto than for the global mean. This means that detecting significant climate change signals (i.e. a change that falls outside the range of historical internal variability) can be more difficult the more localized you get. This is especially true for variables subject to a lot of noise, such as precipitation. Assessing the statistical significance of climate change signals will be discussed further in Chapters 3 and 4, once we start working with climate model data.
2.3.3 Model Structural Uncertainty#
The third major source of uncertainty relates to the different choices in model construction. Differences in the model numerical schemes and in parameterizations mean different models give different answers, and only to a limited extent can we say that one model may be more reliable than another. By comparing model output for a historical simulation to observations, we can quantify the model bias, and one might be tempted to place more confidence in the projections of the model that is the least biased (in other words, the one that best reproduces the observed historical climate). However, there is an issue with climate modeling (and other sorts of modeling of complex systems) called equifinality. Equifinality in this context means that multiple different models can agree equally well with historical observations, but they can each produce different future projections. The model with the lowest bias for the past is not guaranteed to have the best performance in the future. We really have no way to pick a single model to expect to perform best.
Structural uncertainty can be quantified by including a multi-model ensemble in your analysis. The IPCC Assessment Reports frequently use multi-model averages as our best estimates of future climate, though such averaging must be done carefully. The best practice in most cases is to perform analysis for each model individually and aggregate the final results to produce a range of final results that encompass the distribution across models, and thus reflect inter-model uncertainty. Averaging before computing derived quantities can attenuate internal variability and extreme events within individual model simulations, since the sequencing of events in one simulation is completely unrelated to the sequencing of events in another (this also holds true for aggregating data from different ensemble members from a single model). Since different models can produce different projections, but for reasons not well understood, it’s necessary to include multiple models in your climate change impact study to properly quantify the degree of uncertainty in your projections.
Including data from every single available climate model simulation is obviously not feasible for most climate change impact studies. Inevitably you will need to do some model selection and choose a subset of models to include in your full analysis. This may include models that envelope the high and low ends of climate sensitivity, or those which show the largest and smallest projected changes to your main variable of interest. This strategy can allow you to quantify the majority of the possible range of futures while limiting the amount of data you need to process.
Another model selection strategy is to choose the \(N\) models that best match observations for the historical period. This method can lead to undersampling of model uncertainty because of equifinality, but it is still preferable to using only one model in an impact study. Despite being published under different names, and by different modeling centres, many different climate models share common pieces of code. For example, the model E3SM uses the same atmosphere (CAM) and land model (CLM) components as CESM, albeit with some modifications. The Norwegian model, NorESM2 uses these model components as well, and also uses the same sea ice (CICE)) and ice sheet (CISM) models as CESM2. Since the results from these models are not entirely independent of one another, using a validation approach might lead to undersampling of structural uncertainty, since models that share the same code may end up being those which agree best with observations. Selecting a set of models that represent a greater range of underlying methodological choices is important for capturing the true range of model structural uncertainty.
There may also be practical or technical reasons why you may choose to include or exclude certain models from your analysis. One quirk of climate models is that they rarely use the same Gregorian calendar we use in the Western world, where most years have 365 days, except for leap years which have 366. Most model calendars have no leap years at all, and have every simulated year be 365 days long. Others use a 360 day calendar, where each of the 12 months is a uniform 30 days long. Including models with different calendars can cause technical issues when comparing daily climatologies, and it can cause scientific issues when comparing quantities based on annual accumulations, like cooling degree days. To keep things straightforward, you may want to simply exclude models which use a 360 day calendar. More discussion on model selection, and other such study design decisions, will take place in Chapter 5.