It is easy to be a sceptic about climate change modelling: if they can’t get tomorrow’s weather right, what hope is there for a forecast for 50 years ahead? I’d like to examine these statements in order to explain what is involved with modelling climate change and why we should moderate such scepticism. I’ll start with the closely related problem of weather prediction.
The concept of weather prediction based on solving the fundamental equations for the physics of fluid flow was proposed by L.F. Richardson while serving as an ambulance officer during the First World War. He saw that the equations of motion could be solved numerically by dividing a geographical region into a grid and having a ‘computer’ (that is, a person) make calculations for each grid square. Under the instruction of a ‘conductor’ at the centre of a room, each ‘computer’ would pass their results onto their neighbours so that the interactions between grid squares could be taken properly into account.
While his method was somewhat flawed, it did provide the basis for modern numerical weather prediction, and indeed it also anticipated the concept of massively parallel computing which was implemented decades later.
The first useful numerical weather prediction was carried out in the US in 1950 using the first generation of ‘supercomputer’, the ENIAC (Electronic Numerical Integrator and Computer), developed in the Second World War to solve ballistic and other mathematical wartime problems. After the war, John von Neumann, the mathematician who helped develop ENIAC, was looking for peacetime uses of the new computer. Fortunately, at Princeton University he met with Jule Charney, who was one of the leading meteorologists of the 20th century. It was decided that the use of the computer for weather forecasting would be timely and feasible.
After much preparation, the first prediction was carried out in 1950. The calculation was for a 24-hour prediction over North America, and the elapsed time for the calculation was about 24 hours. Although the prediction was not perfect, it demonstrated that numerical weather prediction was possible and most importantly it established a continuing nexus between meteorology and supercomputing.
Today supercomputers are used routinely at a number of centres around the world, including the World Meteorological Centre in Melbourne, to provide the basis for weather forecasting on global, continental and local scales. The useful range of weather prediction has been steadily increasing. However, the inherently chaotic nature of the atmosphere means that there is a limit of about 10 days to deterministic prediction. Beyond this range it is necessary to consider the distribution of probable states of the atmosphere.
Indeed, it is common to run an ensemble of model predictions in order to estimate this probability distribution, even for short-range forecasting. To account for the uncertainty in the initial state of the atmosphere, each ensemble member is started with a slightly different initial condition. The process of assimilating all the routine meteorological observations from around the world (from the ground, ships, buoys, balloons, aircraft and satellites) in order to estimate the initial state of the atmosphere is very complex and consumes a significant amount of supercomputer time.
The consequence of the nexus between supercomputing technology and meteorological science has meant that numerical weather prediction has progressed significantly and the accuracy of forecasts for a few days ahead is adequate to support weather-dependent commercial (for example, aviation, shipping, primary industries) and societal (for example, disaster management, sporting events) activities. It is a very unusual for tomorrow’s weather not to be predicted with useful accuracy.
But increasingly we recognise that the chaotic nature of the atmosphere means that features such as the timing of a front or the intensity of rainfall should be specified in probabilistic terms. These concepts of probable outcomes become vital in understanding climate and climate change modelling.
At the same time as computer models were being developed to predict the weather a few days ahead, scientists were applying the same (or very similar) models to the problem of climate prediction and simulation. A range of questions was being considered; for example, can we simulate the observed seasonal climate variations for various zones of the world or can we predict the mean state of the world’s climate if we changed the concentration of carbon dioxide in the atmosphere?
While weather prediction is essentially an initial value problem in mathematical terms, climate prediction is more of a boundary value problem. That is, weather prediction is very sensitive to the initial state of the atmosphere, while climate prediction is looking more at the impacts of sustained external forcings, such as the incoming solar radiation or greenhouse gases in the atmosphere.
The interpretation of the output of a climate model has to be based on averaging the output over at least time and area; increasingly, ensembles of model results are used to estimate the probable distribution of climate states. It is often difficult to determine how much averaging is needed to obtain a robust estimate of a climate state.
Validation of weather prediction models is relatively straightforward. Under the auspices of the World Meteorological Organization (WMO), national meteorological services routinely (and freely) exchange observed data, which are used to estimate the initial state of the atmosphere to commence a numerical prediction and also to compare with the model predictions for validation purposes. (Both these processes involve sophisticated mathematics because the observations are error-prone and they are often indirect measurements of meteorological variables.) Each day we can make many estimates of the accuracy of the model predictions.
The validation of climate predictions and simulations is much more difficult because we can only consider the results in probabilistic terms. For example, we need to make many year-long runs in order to establish whether a seasonal prediction system is useful for practical applications. On the other hand, because essentially the same model can be used for both weather and climate applications, the continuing validation of a model in weather prediction provides a necessary (but not sufficient) validation for its application to climate problems.
By the late 1980s there were dozens of models around the world being used for climate simulations, and the lack of agreement or validation of these models was limiting the acceptance of the results of any model. Under the auspices of the World Climate Research Programme (WCRP), a program of systematic validation of climate models was established in 1989. It was called the Atmospheric Model Intercomparison Project (AMIP), and it was largely supported by the US Department of Energy.
Any model could be included in the project, provided strict protocols were followed for the initial and boundary conditions of model runs and for the format of model output. Comparisons were carried out between model simulations and the observed climate, and between different models. The AMIP provided credibility to climate models in general, and it also demonstrated that there is no ‘single best’ model: while one model may simulate the best mean climate, another may have a better simulation of the water cycle. The results of AMIP fed into the assessments of climate change science by the Intergovernmental Panel on Climate Change (IPCC), and its approach has been adopted for a range of other model validation projects.
In summary, there have been internationally recognised demonstrations that numerical models can effectively simulate and predict weather and climate variations on a range of scales. However, the chaotic nature of the climate system means that there is increasing uncertainty in predictions as we look further ahead. For this reason we would not attempt a ‘forecast for 50 years ahead’ in the manner of a weather forecast.
On the other hand, we can use our climate models to demonstrate the expected global and large-scale effects of increasing concentrations of greenhouse gases. The chaotic fluctuations due to the internal variability of the climate system act as ‘noise’ masking the greenhouse ‘signal’, but they can be simulated in climate models. Sources of external noise, such as the occurrence of volcanic eruptions that inject ash into the stratosphere or inherent fluctuations in solar radiation, are more difficult to project into the future. However, it has been demonstrated that the effects of these fluctuations can also be simulated in climate models.
We have noted that uncertainty in predictions increases as the prediction period increases. However, uncertainty also grows as we decrease the scale of the region of interest. For example, the range of the annual temperature for a region decreases as we increase the area of the region: the annual average temperature for Melbourne can have a range of about 1.5°C over a decade or so, while the corresponding range for the whole globe is about 0.3°C. It follows that prediction of small-scale climate features has more uncertainty than prediction of large-scale features.
For this reason, analysis of regional impacts of increasing greenhouse gases is usually focused on sensitivity studies that consider the range of possible climate states.
And so, we can answer our sceptic by saying that numerical weather prediction models generally do get tomorrow’s forecast right, and that climate simulations based on these weather models provide useful information on the range of probable climate states decades ahead.
|
