Although weather and climate modeling have common roots in the numerical solution of the governing geophysical fluid dynamics equations of the global atmospheric (initially) and (later) oceanic circulation, the numerical weather prediction (NWP) and climate simulation enterprises have long been entities with different constituencies, distinct detailed technologies, and even specific jargon. Nevertheless, the common origin of these two subdisciplines has been recognized in recent years as fundamental to them both, and efforts to bring them together to address the problem of prediction are now under way (Brunet et al., 2009; Hurrell et al., 2009; Palmer et al., 2008; Shapiro et al., 2010; Shukla et al., 2010; WCRP, 2005). The essence of this “seamless” approach to weather and climate prediction is that they both share common processes and mechanisms, and the interactions across time and space scales are fundamental to the climate system.
Multiple “seamless prediction” strategies are being employed with different aims. They include (i) using suitably initialized IPCC class coupled climate models for hindcasts and predictions on time scales of days to decades; (ii) nesting high-resolution regional models or locally refined grids within global climate models to capture small-scale processes needed to better describe weather events and their statistics; and (iii) using modified versions of operational weather forecasting models for seasonal to decadal prediction. All of these approaches attempt to bridge across the space and time scales of weather and climate.
The ultimate realization of seamless prediction is a single “unified” modeling system designed to work across a broad range of time scales and spatial resolutions, from initialized weather predictions to long-term projections. There are several requirements for a unified weather and climate prediction system:
• data assimilation capability, i.e., to make best use of available observations to generate initial conditions, to facilitate analysis of model parameter sensitivity and uncertainty quantification, and to assess the incremental benefit of new observations;
• unified physical parameterizations and numerical algorithms that can be applied across all scales at which they are needed;
• model forecast skill, evaluated on both weather and intraseasonal to interannual time scales, and model fidelity for long-term climate simulations;
• adequate model development manpower and expertise to simultaneously handle the challenges of both weather and climate applications; and
• computational infrastructure allowing efficient execution and data management for simulations over the needed range of grid spacings and time scales.
Seamless prediction and unified modeling are strategies for better model engineering, aimed at constraining uncertain parameters and taking advantage of the considerable overlap between the internal structure of weather and climate simulation models. Indeed, to the extent such engineering produces more skillful climate models and climate-quality reanalyses, these strategies have clear value to the climate science and applications communities. They may also produce advances that can be transferred between models, including parameterization methodologies that work across a range of scales and new approaches to climate model testing and evaluation.
BENEFITS OF SEAMLESS PREDICTION
The observed climate system contains important features and processes that operate on a wide range of time and space scales, from cloud ice crystals to mesoscale weather systems to basin-scale ocean circulation processes to continental-scale ice sheets. All of these processes contribute to some degree to observed weather and climate phenomena across a range of time scales. Given the complexity of this overall system it has proven useful to construct models that focus on what are deemed the most essential processes for the particular application in mind. For example, weather prediction models used for daily to weekly forecasts focus on high spatial resolution in the atmosphere, and state-of-the-art atmospheric physics, but have less emphasis on a detailed representation of the ocean, because many aspects of ocean changes do not impact the weather on a time scale of a week or two. Similarly, climate models used for projections on decadal to centennial time scales have relatively coarse spatial resolutions for reasons of computational efficiency and thus do not accurately simulate phenomena that may be important on small space and time scales, such as mesoscale convective complexes or tropical cyclones. These choices reflect both limitations on resources, such as computer power, and an attempt to simplify and streamline the problem under consideration.
This approach has its drawbacks, however. Physical and chemical processes in the climate system can have an impact on many time and space scales. For example, small cumulus clouds driven by daytime surface heating can alter afternoon land-surface
temperature and help trigger large thunderstorms (weather effects). They also affect large-scale albedo and surface evaporation (climate effects), so they may play a role in the response of the climate system to changing greenhouse gas concentrations and aerosols on decadal and longer time scales. As a second example, it has become common practice within the past decade for weather prediction models used for forecasting tropical cyclones to employ a dynamical or mixed-layer model of the upper ocean (a traditional climate model component). Such examples have led to increasing recognition that because climate and weather share many of the same underlying physical processes, a more unified approach to model development and application could have many advantages.
For climate models, benefits of a more unified approach include the capability for more rigorous testing and improvement of parameterizations of “fast” physical processes that interact with weather. For example, biases in clouds and uncertainties in the response of clouds to changing greenhouse gases and aerosols are major challenges in projecting climate change over the next century. Biases in clouds appear very rapidly in climate simulations, often within the first days or weeks of a simulation. Therefore, it is appealing to test new parameterizations for clouds in a weather context, where relatively short simulations, initialized from the observed state of the climate system, can provide a rapid assessment of the strengths and weaknesses of model parameterizations.
To this end, one can test weather or climate simulation models in hindcast mode against the large data set of observed past weather variations. Such testing can be done using an initialization from another model. In this report, this testing is referred to as “seamless prediction” but not “unified modeling,” because it does not necessarily require a data assimilation capability that a unified weather-climate forecast model should have to make real-time forecasts. Over the past decade, this approach has started to gain popularity following the development of software infrastructure such as the Climate Change Science Program-Atmospheric Radiation Measurement Parameterization Testbed (Phillips et al., 2004) to support initialization of the atmospheric component of climate models from gridded reanalyses. For instance, Hannay et al. (2009) and Wyant et al. (2010) used global hindcasts by the National Center for Atmospheric Research (NCAR) and Geophysical Fluid Dynamics Laboratory (GFDL) climate models to evaluate their simulation of subtropical boundary layer clouds in specific regions against satellite and in situ observations. The Year of Tropical Convection Madden-Julian Oscillation (MJO) Task Force1 is coordinating multiweek global climate-
model hindcasts of past MJO events. The National Multimodel Ensemble project2 has generated a series of seasonal-interannual hindcasts (and real-time forecasts) by several U.S. and international global climate models.
The unified modeling approach is much more scientifically and organizationally challenging to implement, but it has considerable additional benefits. For weather prediction models, two potential benefits are reduction of systematic errors due to meanstate drift and more skillful data assimilation. Weather forecast models typically suffer from mean-state drifts as they are run out for periods of a few days or longer and drift toward their biased internal climatology, creating forecast errors. In a weather forecast model also designed and tested as a climate model, minimizing such climatological biases will be a development priority; ultimately this should lead to better mediumrange forecasts. Some quantities such as soil moisture affect weather forecasts but are not routinely measured. They therefore are particularly susceptible to large errors due to model drift. Again, model testing in a climate mode should expose such drifts; reducing them can lead to more skillful forecasts and also allow more effective assimilation of observations taken near the land surface, such as near-surface humidity and temperature. A unified model with fewer systematic biases may support more accurate data assimilation and better analyses and reanalyses, which can help in testing of other climate models.
A unified model would foster the development of parameterizations that work well across a range of grid spacings and time scales. Combining weather and climate model resources for development of parameterizations and other modeling infrastructure might ultimately be more efficient and lead to intellectual cross-fertilization between weather and climate model research and development.
Finding 11.1: One useful form of seamless prediction is the testing of climate models in a weather forecast mode. Unified weather-climate modeling has further potential benefits, including improved weather forecasts, data assimilation, and reanalysis, and more efficient use of resources.
SCIENTIFIC CHALLENGES WITH UNIFIED WEATHER-CLIMATE MODELING
This section addresses the scientific and technical challenges for unified modeling, and the management and organizational challenges are discussed later in this chapter.
Using a model for weather prediction requires a methodology for initialization of the model. For real-time forecasting, this typically involves a data assimilation system, which is a major effort, and requires a substantial infrastructure. Thus, one challenge is to optimize the information gained by using a new model to make short-term predictions, while not being overwhelmed with the necessary infrastructure development associated with data assimilation and model initialization.
For weather prediction, detailed analyses of the observed state of the atmosphere are required, but uncertainties in this initial state grow rapidly over several days. Other components of the climate system are typically fixed as observed. For climate predictions, the initial state of the atmosphere makes less difference, but the initial states of other climate system components are necessary. For predictions of a season to a year or so, the upper ocean state, sea-ice extent, soil moisture, snow cover, and state of surface vegetation over land can all be important. For the decadal prediction problem, a full-depth global ocean initial state could be essential (Meehl et al., 2009; Shukla, 2009; Smith et al., 2007; Trenberth, 2008). Initial conditions for the global ocean could conceivably be provided by existing ocean data assimilation exercises. However, hindcast predictions for the 20th century, which are desirable to test models, are severely hampered by poor salinity reconstructions prior to the early 2000s when Argo floats began to provide much better depictions of temperature and salinity in the upper 2,000 m of the near-global ocean. Challenging research tasks are to develop optimal methods for initializing climate model predictions with the current observational network and identifying an optimal set of ocean observations to use for initializing climate predictions (Hurrell et al., 2009).
The mass, extent, thickness, and state of sea ice and snow cover are key climate variables at high latitudes. The states of soil moisture and surface vegetation are especially important in understanding and predicting warm season precipitation and temperature anomalies along with other aspects of the land surface, but they are difficult to quantify. The errors induced by incorrect initial conditions should become less apparent as the simulations evolve as systematic “boundary” and external influences become more important, but they could still be evident through the course of the simulations (Hurrell et al., 2009). Any information on systematic changes to the atmosphere (especially its composition and influences from volcanic eruptions) as well as external forcings, such as from changes in the sun, are also needed; otherwise these are specified as fixed at climatological average values.
Finding 11.2: Current observations are insufficient for complete initialization of climate models, especially for seasonal to decadal forecasts; poorly observed fields will be subject to more initialization bias and uncertainty.
Climate models have coarser grid resolution compared to NWP models, because they must be run for multiyear simulations. A unified model needs to use parameterizations and a dynamical core that can support this range of resolutions, including across “grey zones” where processes such as atmospheric cumulus convection or oceanic mesoscale eddies are simulated but not well resolved. While the European Centre for Medium-Range Weather Forecasts (ECMWF) and the U.K. Met Office (UKMO) have had some success at this, sophisticated theoretical and physical research is needed to try to develop subgrid parameterizations that seamlessly span the NWP-climate range from 10 to 200 km, especially for deep convection. Significant computing resources to facilitate explicit simulation of smaller-scale processes and their interactions with the larger scale will be needed for this research. Its benefits would spread across climate modeling because different applications already use different grid resolutions (e.g., higher-resolution [~50 km] models are used for the decadal prediction problem [e.g., Meehl et al., 2009] for regional climate modeling, or for 10- to 20-year “time slice” global simulations for looking at probability distributions and extremes of weather variability in future climates). Furthermore, within a decade, climate models will use the grid spacings that NWP models use today, so NWP models are a good testbed for developing future climate models.
Finding 11.3: An important challenge for unified modeling (and climate modeling as a whole) is developing improved parameterizations that can work across a range of scales spanning weather and climate applications.
Challenges with model verification are also formidable. Metrics currently used by the climate modeling community differ widely in variable, time scale, space scale, or functional representation. The same is not true in weather prediction, where some estimates of both prediction limits and the impact of different weather prediction metrics can be determined. The skill of daily weather forecasts can be verified many times, and a quantification of model skill is relatively straightforward. The problem is more difficult for seasonal prediction because a large number of seasons and those forecast states must pass in order to build up forecast verification statistics.
For decadal and longer time scales, the problem of quantifying prediction skill becomes even more difficult, and the metrics will likely involve how the forecasts are used in applications. Even if long-term climate models could be tested with all possible climate metrics proposed in the past decade of journal papers, there is no current method to prioritize or weight their impact in measuring uncertainty in predicting future climate change for temperature, precipitation, soil moisture, and other variables of critical interest to society.
Several major numerical weather prediction centers have already spawned unified systems also used for climate modeling. These include UKMO/Hadley Centre, ECMWF, and Meteo-France in Europe, as well as the National Centers for Environmental Prediction (NCEP) (Global Forecast System [GFS]/ Coupled Forecast System [CFS]) in the United States. Initial motivations for developing such systems included seasonal-tointerannual forecasting (NCEP and ECMWF) or an external group interested in developing the weather model into a new climate model (EC-Earth). What can be learned from their experiences?
The most mature unified modeling system is run by UKMO and its climate modeling branch, the Hadley Centre. The Met Office Unified Model, MetUM, was first documented by Cullen (1993) and has been its operational global weather forecast model ever since. It is also the atmospheric component of the HadGEM series of climate models (Collins et al., 2008; Martin et al., 2006). Lastly, regional versions of MetUM are used for high-resolution weather forecasting over the United Kingdom, air-pollution dispersion modeling, and regional climate modeling. Hence, under the unified model umbrella UKMO supports a model hierarchy sharing physical parameterizations, dynamical frameworks, data assimilation, and software infrastructure as appropriate.
Senior et al. (2010) describe the overwhelmingly positive UKMO experience with unified modeling. They note the costs (possible compromises to improved performance on one time scale, additional technical complexity of the modeling and data assimilation system), but they also note “with the MetUM we have encountered relatively few occasions where compromise was required, and more typical is delayed implementation of a change because of lack of performance on a particular timescale.” They stress advantages in rigorous model evaluation, use of diverse observations on many time scales, scale-aware parameterization development, common software infrastructure, and cross-fertilization allowing earlier implementation of new physics (e.g., application of chemical models developed for climate to do air quality forecasting embedded within a weather-prediction model). They also report extensive use of MetUM models by outside academic and applications-oriented user groups.
Martin et al. (2010) show examples of how combined analysis of errors at both fewday and climate time scales stimulated improvements in cumulus parameterization that improved tropical precipitation patterns, and improvements in aerosol and land
albedo parameterization that improved land-surface temperature predictions on both time scales. They report that UKMO is currently working toward full unification of seasonal/decadal climate prediction, which is currently performed with a slightly different modeling configuration, in this framework.
In the past 5 years, UKMO and ECMWF (discussed below) had the highest 5-day weather forecast skill (measured using a standard midlatitude metric, global rootmean-square error in 500 hPa height) of all modeling centers worldwide (Figure 11.1). Both centers have invested heavily in the climate-model strategy of reduction of systematic biases in their forecasts; this aspect of unified model development has clearly been beneficial to their weather forecasts. Gleckler et al. (2008) found that the overall climate simulation biases (averaged over a variety of well-observed global fields) of HadGEM climate model were among the lowest of the CMIP3 climate models, showing the unified strategy also produce a competitive climate model.
FIGURE 11.1 Time series of monthly mean anomaly correlations for 5-day forecasts of 500-hPa heights for GFS/Medium-range Forecast, ECMWF, UKMO, and CDAS (a frozen NCEP model) since 1984, Northern Hemisphere (top) and Southern Hemisphere (bottom). ECMWF has maintained the highest weather forecast skill of all operational modeling centers. SOURCE: http://www.emc.ncep.noaa.gov/gmb/STATS.html/aczhist.html (accessed October 11, 2012).
In 1975, ECMWF was founded by a European consortium to develop a new mediumrange (5-14 days) global weather prediction model, and began operational forecasting in 1979. Since the late 1980s, ECMWF has maintained the highest weather forecast skill of all operational modeling centers (Figure 11.1). In the 1990s, this success led the Max-Plank Institute in Hamburg, Germany, to use a version of the ECMWF model as the basis of a new climate model version, ECHAM4. This was not a true unified model, because a variety of new physical process parameterizations were then added to ECHAM4, and no serious attempt was made to keep ECHAM4 harmonized with further versions of the ECMWF model. However, the ECHAM4 model was among the most skillful of the CMIP3 coupled models at simulating the current climate (Gleckler et al., 2008).
Meanwhile, ECMWF developed its own seasonal-to-interannual forecasting model. In 2005, a new consortium of smaller European climate modeling groups, some university-based, partnered with ECMWF in a project called EC-Earth (Hazeleger et al., 2010), which aims to more fully realize the vision of unified modeling. This project adopted a more intellectually rigorous approach, in which “fast” physics (which respond to atmospheric changes in periods of a few days, such as cloud processes or near-surface soil moisture) are optimized exclusively using weather forecasts (Rodwell and Palmer, 2007). Then long-term climatology and seasonal-to-interannual forecasts are used to optimize “slow” physics such as ocean turbulent mixing or ocean coupling with sea ice that affect the simulation mostly on time scales of months to years, as well as physics that affects both time scales but is most important for climate biases, such as snow density over land. The resulting tuned model slightly improves on the weather forecast skill of the original ECMWF model version and substantially reduces climate biases to a level well below the mean over all CMIP3 coupled models (Hazeleger et al., 2010), showing the value of a seamless approach.
In response to a growing scientific consensus on the potential of coupled modeling for El Niño/Southern Oscillation and seasonal forecasting, NCEP implemented the CFS in 2004 (Saha et al., 2006). The atmospheric model was based on a lower-resolution version of their GFS operational weather forecast system, coupled to externally developed ocean and sea-ice models. Version 2 of the CFS, based on a 2007 version of the GFS with a variety of additional modifications made to improve climate biases and the seasonal forecast skill, became operational in 2011. The CFS can be described
as a “loosely unified” system, in that neither the coupled model development nor the metrics used to assess it currently feed back into changes in the GFS weather-forecast model. A 30-year coupled reanalysis at 50 km horizontal resolution, CFSR, has been performed using CFSv2 (Saha et al., 2010); this is a major additional contribution to climate data that takes advantage of a coupled climate-capable modeling system with cutting-edge assimilation capabilities, and which should help engage the outside community in the CFS effort.
Finding 11.4: Three lessons stand out from examining existing unified modeling systems:
• a unified model can be world leading for both weather and climate simulation;
• successful climate and weather modeling groups that share a unified or near-unified model require a strong supportive management and adequate dedicated resources that can bridge between the different goals and user needs of weather and climate models; and
• unified models are attractive to outside users because of their flexibility and multiscale validation, and help promote interactions between the modeling center and a broad user community whose feedback can improve the model.
THE WAY FORWARD
The committee recommends an accelerated national seamless modeling effort that spans weather to climate time scales. One method to achieve this would be nurturing a U.S. unified weather-climate prediction system capable of state-of-the-art forecasts from days to decades, climate-quality data assimilation, and reanalysis. This prediction system would be a collaboration among operational weather forecast centers, data assimilation centers, climate modeling centers, and the external research community. In particular, it is important to develop it as a partnership between the research and operational communities to best leverage off existing expertise. Versions of this unified model might be deployed as part of an operational prediction system, but it should also be supported for use as a research model.
Ideally, such a model would cross-fertilize parameterization development between the weather and climate communities and naturally lead to parameterization approaches that work well across a range of space and time scales. The committee acknowledges the challenges and risks in such an approach. It requires a clear national-level mandate, strong and skillful leadership, and substantial new resources that recognize that this should be a research effort that can successfully involve a broad scientific community. In particular, these conditions have not been met in the past. No current U.S. modeling center has the resources and capacity to realize this vision on its own. While NCEP’s GFS/CFS has taken important steps toward this unified modeling vision, its further development is subject to both operational constraints and resource limitations. Past experience suggests that partnerships between centers can succeed only if the incentives to work together are strong, sustained, and offer clear scientific opportunities that attract talented scientists and software engineers.
A further management challenge is harmonizing model development for weather versus climate applications. For weather applications, it is advantageous to update the modeling system whenever a proposed change has been demonstrated to improve the forecast skill, because the main application is weather forecasts with a shelf life of a few days. For climate applications, the forecast lead time can be years to decades, and the model output may be bias corrected, downscaled, or used as one step in a chain of models. For such applications, users may prefer a modeling system that remains frozen for several years before an improved version is introduced. Thus, there must be scope for separate development of weather and climate branches of the model, then a periodic, possibly challenging, reintegration of model changes into a single trunk model as at UKMO. This latter step is a defining characteristic of unified model development.
The committee recommends that the U.S. Global Change Research Program, together with the major national climate and weather modeling institutions (e.g., NCEP, GFDL, NCAR, and the Global Modeling and Assimilation Office) work toward defining a unified modeling strategy and initial implementation steps (or deciding this is not a good approach). It should take advantage of the common software infrastructure, community-wide code, and data accessibility. Its success could be judged by simultaneous improvement of forecast and climate simulation skill metrics on all time scales.
One possible benefit of unified modeling is more accurate assimilation of a broad range of observations. Hence, such a unified modeling effort could include research and development of state-of-the-art data assimilation methods, with the goal of pro-
ducing a comprehensive Earth system reanalysis for the past 50 years (or at least for the period 1980 to the present).
Hindcast Testing of U.S. Climate Models
The committee also encourages a nationally coordinated research effort of hindcast testing of all major U.S. climate models (not just the unified model). This is much easier to implement than a unified weather-climate model; each climate model can be run at its preferred grid resolution and need not have a data assimilation capability. The effort could combine several years of hindcasts on weather time scales (up to 15 days) and coupled-model hindcasts on intraseasonal to interannual time scales. Each model could use either externally initialized fields or some form of relaxation or data assimilation. A rigorous and coordinated testing process using a standardized protocol, outputs, and diagnostics would facilitate model intercomparisons and accelerate progress. Tests could include perturbed initial conditions, perturbed-parameter ensemble hindcasting capability, and perhaps ensemble Kalman filter data assimilation to guide the choice of “fast physics” parameters. Results should be made publicly available in a standard web-accessible form. The main goals would be to evaluate and improve model representations of “fast” physical processes that vary strongly on these time scales and to optimize uncertain parameters within these representations.
Recommendation 11.1: To fully exploit a multiscale approach to model advancement, the United States should nurture a unified weather-climate prediction system capable of state-of-the-art forecasts from days to decades, climate-quality data assimilation, and Earth-system reanalysis.
Recommendation 11.2: To reduce sources of uncertainty in climate simulations, the United States should pursue a coordinated research effort to use weather and/or seasonal/interannual hindcast simulations to systematically constrain uncertain parameters and to improve parameterizations in its major climate models.