gfdl's home page > about us > science of GFDL > science and nonscience concerning human-caused climate warming > why climate models are imperfect and why they are crucial anyway
why climate models are imperfect and why they are crucial anyway
Over the past three decades, a quiet revolution has fundamentally changed the way that much of the research in climate science works. Earlier, the controlling science paradigm was the interchange between theory and observation concerning the structure and behavior of natural phenomena. Today, much climate research is driven by the interactions among theory, observation, and modeling. By modeling, we mean computer-based simulations of various phenomena based on numerical solutions of the theory-based equations governing the phenomena under investigation. These combined approaches are now widespread in the physical sciences. It is significant that mathematical modeling of weather and climate literally pioneered this new approach to scientific research.
Mathematical models of climate can range from simple descriptions of simple processes to full-blown simulations of the astoundingly complex climate system. Models of the coupled atmosphere-ocean-ice-land system lie close to the most complex limit of such models. This very complexity of climate models can lead to highly divergent human reactions to them, varying from "garbage in, garbage out" to almost worshipful. The truth is far from either of these unscientific characterizations.
Newcomers to the greenhouse warming problem tend to be unaware of the long and rich history of mathematical modeling of the atmosphere and the ocean. In the late 1940s and early 1950s, simple mathematical models were created to attack the weather forecasting problem. More advanced models were built in the late 1950s and early 1960s (6, 7) because of a strong research interest in understanding the circulation of the atmosphere. Shortly thereafter, the first model bearing a strong resemblance to today's atmospheric models was created (8). That early model, as well as all of today's models, solves the equations of classical physics relevant for the atmosphere, ice, ocean, and land surface. These equations are conservation of momentum (Newton's second law of motion), conservation of heat (first lawof thermodynamics), and conservation of matter (air, water, chemicals, etc, can be blown around by wind or currents, changed in phase, transferred across boundaries, or converted chemically, but the number of atoms of each kind remains unchanged).
The modeling approach thus provides high potential for fundamental tests of applications of these theoretical first principles. Such modeling appears deceptively simple: These equations are taught in high school physics. There are some daunting challenges, however. When coupled and applied to moving (and deforming) fluids such as air and water, these equations form continuum systems that are intrinsically nonlinear and can exhibit surprisingly counterintuitive behaviors. Moreover, their solution in a climate model requires a reasonably fine-scale grid of computational points all over the atmosphere-ice-ocean-land surface system. In addition, important small-scale processes such as moist convection (e. g. thunderstorms) and turbulent dissipation remain formidably difficult to incorporate on a first-principles basis. Worse, no meaningful steady-state solutions solve directly for the average climate. In effect, the average climate in such a model must be described as a statistical equilibrium state of an un-stable system that exhibits important natural variability on timescales of hours (thunderstorms), days (weather systems), weeks to months (planetary-scale waves/ jet-stream meanders), years (El Niño), and decades to centuries (ocean circulation variations and glacial ice changes). Clearly, models of such a large and complex system are intrinsically computer intensive. Fortunately, today's supercomputers are over a thousand times faster than those of 30 years ago. Because of today's widespread availability of relatively inexpensive computer power, the number of fully coupled atmosphere-ocean climate models in the world has increased from a few in the early 1980s to roughly 10 independently conceived models today. Roughly 20 more are essentially based on these 10 models.
Over the last half century, use of these kinds of physically based mathematical models has resulted in major improvements in the science of weather forecasting. Sharp skill improvements have been achieved in finding the useful short-term predictability in a fundamentally chaotic system (by which I mean that the details of weather variations become essentially unpredictable after a sufficient lapse of time, say a couple of weeks) (9). For example, it has become almost routine to forecast the intensity and path of a major winter storm system well before the surface low-pressure area (so ubiquitously displayed in television weathercasts) has even formed.
Recently, it has become clear that slower variations of the coupled ocean-ice-atmosphere- land surface system provide potential for finding useful predictability on timescales longer than the couple of weeks characteristic of individual weather systems. The most visible example is the realization that El Niño events, which produce warming in the tropical eastern Pacific Ocean, may be predictable a year or so in advance under certain circumstances (10). The existence of such a "predictable spot" of warm ocean suggests a "second-hand" improvement of prediction of seasonal weather anomalies (e. g. a wetter-than-normal California winter).
The existence of such extended-range predictive potential in the climate system leads to obvious questions about such models' validity for predicting systematic changes in the statistical equilibrium climate (say a 20-year running average) resulting from the inexorably increasing infrared-active gases that are currently underway. First, we must recognize that these are conceptually quite different things: Weather forecasting attempts to trace and predict specific disturbances in an unstable environment; climate projections attempt to calculate the changed statistical equilibrium climate that results from applying a new heating mechanism (e. g. CO2 infrared absorption) to the system. Perhaps surprisingly, predicting the latter is in many respects simpler than predicting the former.
As an example of the fundamental difference between weather forecasting and climate change, consider the following simple and do-able "lab" thought experiment that utilizes the common pinball machine.8 As the ejected ball in the pinball machine careens through its obstacle-laden path toward its inevitable demise in the gutter, its detailed path, after a couple of collisions with the bumpers, becomes deterministically unpredictable. Think of this behavior as the "weather" of the pinball machine. Of course, the odds against success can be changed dramatically in favor of the player by raising the level of the machine at the gutter end, in effect changing the "climate" of the pinball machine. By reducing the slope of the playing field, the effective acceleration of gravity has been reduced, increasing the number of point-scoring collisions before the still inevitable final victory of gravity. Interestingly, in this altered pinball machine "climate," the individual trajectories of the balls are ultimately as unpredictable as they were in the unaltered version. The diagnostic signal of an altered pinball "climate" is a highly significant increase in the number of free games awarded. A secondary diagnostic signal, of course, is a noticeable decrease in the received revenues from the machine. It thus is conceptually easy to change the pinball machine's "climate." Detecting changes in pinball machine "climate" and attributing its causes, however, can be easily obscured by the largely random statistics of a fundamentally chaotic system, not unlike in the actual climate.
What do these pinball machine experiments have to do with understanding models of the real climate? Projections for greenhouse warming scenarios depend on a number of physical processes (see above) that are subtle, complex, and not important to weather prediction. However, people outside the climate field are frequently heard to say that climate models are ill posed and irrelevant because they attempt to forecast climate behavior that is well beyond the limits of deterministic predictability and that if one cannot predict weather more than a week in advance, the climate change problem is impossible. Such statements are scientifically incorrect. The "weather prediction" problem is essentially an initial value problem in which the predictability of interesting details (i. e. weather) is fundamentally limited by uncertainty in initial conditions, model errors, and instabilities in the atmosphere itself. In contrast, climate change projections are actually boundary value problems, (e. g. interference with a pinball machine's acceleration of gravity), where the objective is to determine the changes in average conditions (including the average features of the evolution toward the new equilibrium) as the planet is heated or cooled by newly added processes (e. g. increased CO2 ).
The differences between weather and climate models are further instructive when one considers how their strengths and weaknesses are evaluated. Thanks to massive amounts of weather and climate data, both kinds of models can be evaluated by careful comparison with data from the real world. In practice, however, the approaches to improving these superficially similar models are very different. The weather models are evaluated by comparing model-based forecasts, started up from real data on a given day, with what happened hours to weeks later. Interestingly, one of the key problems with such weather models is that they can easily reject their initial conditions by drifting toward a model climate that is quite different from that of the real data that was used to start up the detailed forecast calculation. In effect, such a weather forecast model is deficient in the climate that it would produce if released from the constraints of its starting data.
In sharp contrast, a climate model has the responsibility of simulating the time-averaged climate for, say, today's conditions (or for around, say, the year 1800). In this case, the focus of the scientific inquiry is quite different. Here, attention is directed toward proper simulation of the statistics of climate, such as the daily and annual temperature cycles forced by the sun, the number and intensity of extratropical cyclones, locations of deserts and rainy areas, strength and location of jet streams and planetary waves, fidelity of El Niño simulation, location and characteristics of clouds and water vapor, strength and location of ocean currents, magnitude and location of snow accumulation and snow melt, and, finally, amplitudes and patterns of natural variability of all of these on a wide range of timescales (days to centuries).
Achieving all of this in a climate model is a daunting task because the enormous wealth of phenomena in the climate system virtually requires the use of judicious tuning and/ or adjustment of various poorly defined processes (such as clouds, or the fluxes of heat between atmosphere and ocean) to improve the model's agreement with observed climate statistics. Such tunings and adjustments are widespread, especially for the global-mean radiative balance, and are often done to ensure that the model agrees with the global-mean features of the climate. If this is not done, a coupled model started up with today's climate will tend to drift toward a less realistic climate. These practices have been criticized as evidence that climate models have no credibility for addressing the greenhouse warming problem. Interestingly, such tunings and adjustments (or lack thereof) may have little to do with the ability of a model to reduce its fundamental uncertainty in predicting anthropogenic climate change. Recall that the key uncertainties highlighted above (water vapor, cloud, and ice albedo feedbacks) revolve around how such properties might change under added greenhouse gases. This is a set of modeling problems that cannot be evaded by judicious model tuning or adjustments. Likely to prove much more fruitful in the long run would be improved fundamental modeling of the key processes that govern the most important climate feedback processes as CO2 increases (e. g. clouds, water vapor, ice, ocean circulation).
Thus, the models are imperfect tools with which to make such climate-change predictions. Does this mean we should shift our focus to other tools? Definitely not. Statistically based models that use historical data are possible alternatives, but they are of marginal validity, mainly because the recent earth has never experienced the rate of warming expected to result from the current runup of infrared-active greenhouse gases. In this sense, the large, but very slow, global-mean climate excursions of the past geological epochs are instructive, but they are far from definitive as guidelines or analogs for the next century.
The above considerations make it clear that there is no viable alternative to coupled climate models for projecting future climate states and how they might unfold. The physically based climate models have the huge advantage of being fundamentally grounded in known theory as evaluated against all available observations. There are indeed reasons to be skeptical of the ability of such models to make quantitatively accurate projections of the future climate states that will result from various added greenhouse gas scenarios. Fortunately, the weak points of such climate models can be analyzed, evaluated, and improved with properly focused, process-oriented measurements, complemented by well-posed numerical experiments with various formulations of the climate models.9 In short, the use of such climate models allows a systematic approach to close the gap between theory and observations of the climate system. No alternative approach comes close.
8 The pinball machine is a device designed for recreation and amusement that allows the player to shoot steel balls (of roughly 1-in diameter) into an obstacle-strewn field of electronic bumpers that, when struck by the ball, act to increase the net speed of the ball (super elastic rebound). The playing field is slanted so that the ball enters at the highest point. When all five balls have been trapped in the gutter, the game is over. The object of the game is to keep the balls in play as long as possible (through adroit use of flippers near the gutter that propel the ball back uphill and away from the dreaded gutter). The longer the ball is in play, the more it is in contact with bumper collisions that add to the number of points earned. A sufficiently high score wins free replays. Thus, the object of the game is for the player's skill to overcome gravity for as long as possible, somewhat analogous to the efforts of ski jumpers and pole vaulters.
9Out of many
such examples, one of the more interesting is provided by the Department
of Energy's Atmospheric Radiation Measurements Program. At a heavily instrumented
site in Oklahoma (and at some lesser sites), intensive measurements are
made of horizontal wind, vertical velocity, temperature, water vapor, clouds,
latent heating, precipitation, short-and long-wave radiative fluxes, and
surface fluxes of heat, momentum, and water vapor. This comprehensive set
of measurements is being used to evaluate our current modeling capabilities
and deficiencies on cloud processes, "cloudy" radiative transfer,
convection (thunderstorm scale), and turbulence. These areas represent
some of the weakest aspects of the atmospheric parts of climate models.
Next: Why Climate Data are Imperfect and Why They are Crucial Anyway