Seamless Prediction Systems by Hendrik Tennekes

Guest Weblog by Henrick Tennekes  June 24 2008

Roger Pielke gracefully invited me to write a brief essay on an interesting technical detail in the World Summit document issued by WCRP. According to the document, all time scales, from hours to centuries, all regional details, everything related to prediction should be dealt with by GCM technology. In this context, the term “seamless prediction” is used. That caught my attention. Let me quote the relevant paragraph:

“Advances in climate prediction will require close collaboration between the weather and climate prediction research communities. It is essential that decadal and multi-decadal climate prediction models accurately simulate the key modes of natural variability on the seasonal and sub-seasonal time scales. Climate models will need to be tested in sub-seasonal and multi-seasonal prediction mode also including use of the existing and improved data assimilation and ensemble prediction systems. This synergy between the weather and climate prediction efforts will motivate further the development of seamless prediction systems.”

The current use of the concept of seamless prediction is explained in a recent paper by Tim Palmer and others, published in the Bulletin of the AMS (see Palmer, T.N., F.J. Doblas-Reyes, A. Weisheimer, and M.J. Rodwell, 2008: Toward Seamless Prediction: Calibration of Climate Change Projections Using Seasonal Forecasts. Bull. Amer. Meteor. Soc., 89, 459–470. ). I quote:

“If essentially the same ensemble forecasting system can be validated probabilistically on time scales where validation data exist, that is, on daily, seasonal, and (to some extent) decadal time scales, then we can modify the climate change probabilities objectively using probabilistic forecast scores on these shorter time scales.”

“We propose that if the same multimodel ensemble is used for seasonal prediction as for climate change prediction, then the validation of probabilistic forecasts on the shorter time scale can be used to improve the trustworthiness of probabilistic predictions on the longer time scale. This improvement would come from assessing processes in common to both the seasonal forecast and climate projection time scales, such as the atmospheric response to sea surface temperatures. To reiterate, our basic premise is that processes, such as air–sea coupling, that are relevant for the seasonal forecast problem also play a role in determining the impact of some given climate forcing, on the climate system itself. The calibration technique provides a way of quantifying the weakness in those links to the chain common to both seasonal forecasting and climate change time scales.”

Apparently, the idea behind this application of the seamless prediction paradigm is that the reliability of climate models can be improved if they are used as extended-range weather forecast models. Experimental verification, which is impossible in climate runs, then becomes feasible. With a bit of luck, certain types of shortcomings in the model formulation can be detected this way. This process may lead to climate codes with fewer systematic errors.

This sounds promising. Climate models cannot be verified or falsified (if at all, because they are so complex) until after the fact. Strictly speaking, they cannot be considered to be legitimate scientific products. Any methodology that would ameliorate this situation would be a step forward, however small and tentative. I am happy to grant Palmer et al the benefit of the doubt as far as this point is concerned.

But I wonder how short-term calibration of a long-term tool might help to unravel the long-period irregularities in the climate system. The original meaning of the term “seamless prediction” was to express the idea that weather forecasting technology can be usefully extended to climate problems. The term was coined to consolidate the monopoly of GCM technology in all kinds of weather and climate forecasting. However, in the paper by Palmer et al. it refers to the reverse focus, where calibration is attempted by shrinking the time horizon. Alice gazing through the other side of the looking glass, as it were.

The tail wags the dog here. I know that dressed-up versions of weather forecast models are used to make climate prediction runs. I don’t mind too much, though this methodology hides a chronic, distressing lack of insight in the statistical dynamics of the General Circulation. I consider the seamless use of GCM technology a sign of intellectual poverty. Gone are the days of Jule Charney’s Geostrophic Turbulence, Ed Lorenz’ WMO monograph on the General Circulation, and Victor Starr’s early thoughts on Negative Eddy Viscosity Phenomena.

To turn the matter on its end is one step too far. Short- and medium-term forecast methods work quite well without an interactive ocean, interactive biosphere, interactive changes in the state of the world economy, and the like. I see no reason to burden a weather forecast model with the enormous complexity of climate models, and I see no way in which interactions of subordinate importance in weather forecasting can reliably be calibrated to improve crucial interactions in climate runs. I know I rub against the grain of the GCM paradigm, but so be it.

Palmer et al. also seem to forget that, though weather forecasting is focused on the rapid succession of atmospheric events, climate forecasting has to focus on the slow evolution of the circulation in the world ocean and slow changes in land use and natural vegetation. In the evolution of the Slow Manifold (to borrow a term coined by Ed Lorenz) the atmosphere acts primarily as stochastic high-frequency noise. If I were still young, I would attempt to build a conceptual climate model based on a deterministic representation of the world ocean and a stochastic representation of synoptic activity in the atmosphere.

One example I am familiar with is the North Atlantic storm track, which guides the surface winds that drive the Gulf Stream and help to sustain the thermohaline circulation in the world ocean. The kind of model I envisage deals with the slow evolution of the ocean circulation deterministically, but with the convergence of the meridional flux of atmospheric eddy momentum in the way turbulence modellers do. In this view, the individual extra-tropical cyclones that feed the momentum of the jet stream can be represented by stochastic parameterizations, but the jet stream itself is part of the deterministic code. In a more general sense, I claim that stochastic tools of the kind proposed by Palmer et al. will have to be developed on the basis of a better understanding of the dynamics of the climate system. Purely statistical methods, however sophisticated, can be compared with attempts to kill a songbird with a shotgun.

There is yet another principal shortcoming in the paper by Palmer et al. I will grant them that the approach they advocate may be of some use as far as the possible deleterious effects of greenhouse gases are concerned. These gases are rapidly mixed through the entire atmosphere. That’s what the turbulence in the general circulation is good at. But now think of slow forestation and deforestation, or the expected northward crawl of corn and wheat belts.  And what about large hydropower projects or land-use changes as the peoples of India and China become wealthier, drive more cars, and become more urbanized? Can the reverse use of seamless prediction methods help to calibrate the response of the climate system to these elements of the Slow Manifold? I would not know how.

I offer a solution to Palmer’s quandary. Seamless prediction may or may not have a glorious future, but it does have a history spanning almost twenty years. I propose that WCRP should initiate a Seamless Reprediction Program, as a kind of extension to the reanalysis efforts undertaken from time to time at ECMWF. That is, climate runs made in the past should be analyzed, restarted with the latest version of the stochastic feedback paradigm, and calibrated with accumulated observational evidence. Perhaps the latest versions of climate models cannot be investigated this way, but the great advantage is that working in a retrospective mode offers falsification prospects. Looking back, all data needed for calibration do exist. So do the computers and the software. Immediate, large-scale expansion of facilities is not needed if this path is taken. And I trust ECMWF will be permitted to participate in this effort.

Would Palmer not agree that evidence from such a Reprediction Program might turn out to become a cornerstone for the World Climate Computing Facility that he and the World Summit crowd are lobbying for? I wish them well.

Comments Off on Seamless Prediction Systems by Hendrik Tennekes

Filed under Guest Weblogs

Comments are closed.