| Developing Models for Predictive Climate Science |
The Community Climate System Model results from a multi-agency collaboration designed to construct cutting-edge climate science simulation models for a broad research community. Predictive climate simulations are currently being prepared for the petascale computers of the near future. Modeling capabilities are continuously being improved in order to provide better answers to critical questions about Earth’s climate. |
| Climate change and its implications are front page news in today’s world. Could global warming be responsible for the July 2006 heat waves in Europe and the United States? Should more resources be devoted to preparing for an increase in the frequency of strong tropical storms and hurricanes like Katrina? Will coastal cities be flooded due to a rise in sea level? The National Climatic Data Center (NCDC), which archives all weather data for the nation, reports that global surface temperatures have increased over the last century, and that the rate of increase is three times greater since 1976. Will temperatures continue to climb at this rate, will they decline again, or will the rate of increase become even steeper? |
| To address such a flurry of questions, scientists must adopt a systematic approach and develop a predictive framework. With responsibility for advising on energy and technology strategies, the DOE is dedicated to advancing climate research in order to elucidate the causes of climate change, including the role of carbon loading from fossil fuel use. Thus, climate science—which by nature involves advanced computing technology and methods—has been the focus of a number of DOE’s SciDAC research projects. |
Dr. John Drake (ORNL) and Dr. Philip Jones (LANL) served as principal investigators on the SciDAC project, “Collaborative Design and Development of the Community Climate System Model for Terascale Computers.” The Community Climate System Model (CCSM) is a fully-coupled global system that provides state-of-the-art computer simulations of the Earth’s past, present, and future climate states. The collaborative SciDAC team—including over a dozen researchers at institutions around the country—developed, validated, documented, and optimized the performance of CCSM using the latest software engineering approaches, computational technology, and scientific knowledge. Many of the factors that must be accounted for in a comprehensive model of the climate system are illustrated in figure 1.
|
The Climate System
The first theoretical construct required for the study of climate is the principle of conservation of energy. Any energy that enters the system must eventually be removed for the principle of conservation to hold. This applies equally at various scales—from a small patch of ground, an entire ocean layer, or an astronaut’s blue marble view of the Earth.
|
| The Earth is a thermodynamic system, and climate serves as the energy flow process required to maintain equilibrium. The climate system achieves this stability by balancing energy input from the Sun with energy output via radiation from the planet’s surface. For example, one job of the climate system is to transfer incoming energy from the sun-soaked tropics to the cold higher latitudes, where heat energy is radiated back out into space. |
 |
Figure 1. The components of the Earth's climate system first interact vertically to establish a balance between the incoming solar energy and the outgoing heat energy that is radiated to space. The reflectivity of the Earth’s surface (albedo) depends on vegetation type and whether the surface is land, lake, or ocean. The moisture and chemical content of clouds also change the reflectivity of cloud tops and the absorption spectrum for different wavelengths of light. The horizontal arrows in the diagram represent interactions involving the transport of material by winds, currents and runoff. The atmospheric general circulation interacts with the ocean general circulation, each driving the other on a global scale.
|
|
Components of the Earth’s Climate
Wind patterns in the atmosphere such as Hadley cells (see sidebar "Wind Belts and Circulation Cells") and ocean currents like the Gulf Stream are a combined product of the Earth’s rotation and the unequal heating of the planet’s surface. Winds and currents are the mechanisms that redistribute energy around the planet. These redistributions must abide by the physical laws of conservation of mass and momentum. For example, the chemical balance of gases in the atmosphere determines that incoming shortwave sunlight is absorbed and reflected longwave heat is trapped. The composition of gases and particulates in the atmosphere—including water vapor, carbon dioxide, methane, ozone, and aerosols—is determined by factors at the surface of the Earth, such as plant growth, soil properties, and human activities. Because of its reflectivity, ice also affects solar absorption and is an important feedback element in the climate system. The complexity of the interrelated factors is demonstrated in figure 1, which illustrates mass and energy exchanges among various climate system components. |
Modern day climate models such as the third-generation Community Climate System Model (CCSM3) include the components shown in figure 1 into a detailed, first principles theoretical framework. SciDAC researchers supported by the collaborative endeavor of DOE, NSF, and NASA have worked to integrate this framework into computer code for use on powerful supercomputers. The CCSM project has bridged the gaps between disciplines, applying computer science to rigorous theory in order to develop practical scientific simulations.
|
Modeling Climate in Detail
Climate is often thought of as a weakly forced and strongly nonlinear dynamical system (sidebar "Dynamical Systems," p48). The major factor forcing the climate system is the solar constant (1,367 W/m2). Solar input variation—due to the sunspot cycle and the seasonal distance between the Earth and the Sun—are small, amounting to less than 4 W/m2. Increases in greenhouse gases in the atmosphere change the longwave absorption by an equivalent amount. The inherent nonlinearities of the system are evident in the constantly shifting weather patterns that arise from differential heating of the Earth’s surface and instabilities of atmospheric flow. In terms of creating a reliable model, this means that fundamental physical laws must be treated carefully. Furthermore, factors such as process details, topography, bathymetry, and land use must also be considered. The resolution of the model must adequately represent all of these relevant dynamical interactions.
|
| The nature of the climate problem and its dependence on what John Von Neumann called the “minutiae of computation” make qualitative arguments ineffective for developing predictive capability. The range of possible climate states and the variability of the Earth’s weather are the result of rather complex feedback loops and balances among the system components. |
| As with many scientific endeavors that are simply not practical for the laboratory bench, computational experiments have emerged as the only reliable tools for developing verifiable, quantitative predictions. A reasonable explanation of the behavior of the climate system can result when simulation experiments are combined with careful validation and verification using observational data, such as from satellites and other tracking systems. However, without the theoretical basis expressed in the modeling framework, it would be impossible to develop predictive capability from such data alone. |
 |
Figure 2. Shown here are the prevailing wind belts and primary circulation cells of the climate system on Earth. Patterns similar to those in the Northern Hemisphere also occur in the Southern Hemisphere.
|
|
Computational Methods
The development of new climate models is progressing at a rapid pace. The National Center for Atmospheric Research (NCAR) has been a major player in model development. At NCAR the component models have multiplied and the physical principles they represent have increased in complexity.
|
| Resolution for atmospheric calculations has doubled twice in the last five years, with grid spacing reduced from 2.8° (300 km) to 1.4° (150 km), and then down to 0.7° (75 km). This progression continues to benefit from weather forecast improvements at the European Center for Medium Range Weather Forecasting and other centers, where information-rich operational resolutions of 0.175° (20 km) are currently in use. |
| Like most scientific simulation projects, modeling the climate system is constrained by finite computing time resources. Thus, less informative resolutions are used for climate models than for weather models. In practice this amounts to trading some spatial accuracy for the ability to make projections on very long time scales. For example, weather forecasting involves trends specific to small spaces over short periods of time, such as the next few days or weeks. Climate modeling, on the other hand, deals with large trends over time periods of decades, centuries, or longer. Hence, some spatial accuracy can be sacrificed in order to make long-term climate simulations computationally practical. |
| Of course, this tradeoff inhibits the ability to provide information for regional climate projections, and so there is always a demand for increased resolution in climate simulations. In fact, many physical processes are inadequately represented at current resolutions. In one such example of using higher resolution, simulations using the Parallel Ocean Program (POP2) have been carried out by SciDAC investigators and their collaborators at resolutions of 0.1° and greater. This permits simulated ocean eddies to form and transport heat in much the same way as observed in the actual ocean. |
| Additionally, a new algorithm for atmospheric flow is experiencing wider use within the modeling community. The new method involves a flux-form semi-Lagrangian formulation and a finite-volume discretization of the spatial extent. This algorithm replaces the Eulerian and semi-Lagrangian spectral dynamical cores of the Community Atmospheric Model that were used for the Intergovernmental Panel on Climate Change (IPCC) simulations (sidebar "Intergovernmental Panel on Climate Change," p49). The improved version uses a special technique called the Piecewise Parabolic Method (PPM) to construct fluxes in a vertical coordinate. Since this method is strictly conservative, it has been the method of choice for an increasing number of scientists interested in atmospheric chemistry, aerosols, ozone, and carbon dioxide distributions. |
Efficient implementation of traditional methods as well as novel numerical techniques has dramatically boosted the productivity of scientists using the CCSM and its component models. Programming practices adopted from software engineering have led to logistical improvements—making the models maintainable, easily extensible, and portable across a wide variety of computing equipment while maintaining high computational performance.
|
Planning for Petascale Computing
The challenge represented by the exponential growth in computing capability is entering a new and exciting chapter. Scientists are preparing to utilize a computer—planned to be available in 2009—with 25,000 to 100,000 processors delivering a petaflop of computational power.
|
| In theory, petascale machines can perform 1,015 floating point operations per second, but software tools run on such machines must be designed specifically to exploit the awesome power of the hardware. The current approach to parallel computing in the CCSM is based on two principles: (1) decompose the data structure of each component to provide parallel throughput, and (2) remap fields from one data structure to another as needed. Until recently little work has been done on parallel algorithm design factors, such as maintaining good load balance, scaling to large processor counts, or minimizing overall communication costs. Generally, the parallel implementations of the past few years have fit well with available supercomputers, which are based on powerful vector or server architectures tightly linked with high speed networks. Unlike these machines, the next generation of petascale computers will be designed to utilize a staggering number of relatively weak processors coupled together with slow networks. Computers of this architecture are more sensitive to task scheduling, load imbalance, and hardware failure, and thus require new approaches and careful software design. |
| SciDAC-sponsored computer science research has led to tools and techniques critical to tackling the petascale challenge. In order to deal with larger quantities of output, the Scientific Data Management project (SciDAC Review, Fall 2006, p28) generated a powerful input/output tool called Parallel NetCDF. New software component technologies for coupling parts of a computation—like the Common Component Architecture (CCA) project—as well as performance monitoring tools are available for measuring how each processor is coping with the load assigned to it. Development continues on new system tools for scalable operating systems. New mathematical techniques from the Terascale Optimal PDE Simulation (TOPS) project (SciDAC Review, Spring 2006, p50), such as the Newton–Krylov methods and the discontinuous Galerkin discretizations, are also being prototyped with new grid systems from the Terascale Simulation Tools and Technologies (TSTT) project. These advances are being set up for use in atmospheric and ocean flow problems. Overall, the SciDAC program has spawned a rich network of expertise and an infrastructure of methods and technology that will help advance climate modeling to the new levels made possible by petascale computing. |
It might seem logical that increasing the number of available processors by a factor of ten equates to running ten times as many jobs simultaneously. Although this obvious parallelism would help improve the statistical information from ensemble forecasts, current work is focused on exploiting the increased processing power for more comprehensive modeling of the climate. The computational cost of the simulation grows as additional features—such as robust atmospheric chemistry, ocean ecosystems, and dynamic land vegetation—are added to the model. A coupled global Earth system model would simulate important factors such as changing soil and ocean carbon pools that may influence the long-term balance of carbon dioxide in the atmosphere. An increasing focus on the regional impact of climate change will also require higher resolution models. While this is an obvious scaling opportunity for new architectures, new computational methods will be required to address the resulting time-step constraints for century-scale simulation. The new DOE computer platforms and SciDAC algorithm development will make such advances possible as the challenge shifts to understanding and utilizing new climate simulation results.
|
Results—What the Models Predict
Climate involves weather averaged over time combined with statistics accounting for the frequency of extreme conditions. It makes sense to talk about the climate of the mid-century—a broad trend which we have some hope of being able to predict—but it is impossible to forecast the exact weather in 2056.
|
| A standard exercise in climate prediction is to project possible outcomes for the last decade of the 21st century based on projected greenhouse gas emissions. For a medium range emission scenario, the CCSM3 predicts a warming of 1.1 °C by mid-century and 2.3 °C by the end of the 21st century. The warming is accompanied by other results. A 25 cm rise in sea level has been predicted. The North Atlantic conveyor belt-like circulation (sidebar "Ocean Modeling and the Interaction of Ocean Ecosystems," p50) will slow by about 24%, but the model indicates that, if carbon dioxide levels are stabilized, the circulation pattern will recover over the following two centuries. Most temperature changes will occur in the high latitudes near the poles while the mid-latitudes will likely experience more moderate temperature increases. An El Niño-like response to the warming will provoke increased precipitation in the tropics and mid-latitudes but a decrease in the subtropics. Precipitation from monsoon rains will increase. The model results show no significant soil drying in the Northern Hemisphere mid-latitudes. Rainfall will typically be heavier throughout the year but this factor will be offset by reduced winter snowfall. Warmer temperatures should lead to greater evaporation from the soil, and thus late summer soils will tend to be drier at the end of the 21st century, relative to the present. |
 |
| Figure 4. Taylor diagrams for the finite-volume General Circulation Model (fvGCM) model (the arrows go from the 2 x 2.5 to the 0.4 x 0.5 resolution version) compared to observations over the period 1979 to 1999. This test is defined as part of the Atmospheric Model Intercomparison Project (AMIP) sponsored by the DOE Climate Change Prediction Program. This study information was provided by Dr. Phil Duffy of LLNL. |
|
A key unknown deals with what will happen to the polar caps and ice sheets. Model predictions indicate that the caps and ice sheets will become increasingly vulnerable to melting from below during the summer as the North Atlantic sea temperatures rise and the circulation pattern slows and lengthens. Melting ice sheets are only now being added to models as observations indicate the Greenland and West Antarctic ice sheets may be melting more rapidly than previously realized. Adding important factors like this melting phenomenon is an example of why the models must be continuously updated and prepared for simulations on more powerful computers.
|
Verifying Models—Confidence Levels
To determine the validity of new climate models, two simulations are subjected to close and careful scrutiny. The first verification test is a control simulation that fixes the boundary conditions and runs for several centuries. This simulation tests the equilibrium of the model and yields information concerning the “drift” and stability of the model, It also checks the balance of processes, like pole-ward heat transport, and provides a comparison with observed climatology.
|
| The second simulation test compares historical climate data with the simulated climate generated by the model. The model is forced in a time-dependent fashion with actual atmospheric carbon dioxide concentrations, solar variability, and dust from volcanic eruptions. Satellite and weather station data are used to identify where improvements need to be made. |
| The moment of truth in any modeling project comes when the final version of the model is compared with observations. Figure 4 depicts a Taylor diagram, a specialized graph used for communicating these comparisons between model and empirical data. More than twenty output fields are measured and correlated with observational data using standard deviation as a measure. If an arrow points toward zero on the Taylor diagram, the deviation of the output has been reduced and the correlation has improved, that is increased from one version of the model to the next. The diagram in figure 4 compares a low resolution model to a high resolution version of the Community Atmosphere Model (CAM) with the finite volume dynamical core. It can be concluded that increasing the resolution resulted in significantly better predictions for some fields while only marginally improving others. |
 |
Figure 5. A snapshot of the simulated time evolution of the component of atmospheric carbon dioxide (CO2) concentration originating from the land surface for February 1900. This “green CO2” is a product of the net ecosystem exchange (NEE) and the CO2 flux due to respiration of vegetation (autotrophs) and soil microbes (heterotrophs) minus that taken up for ecosystem production. CO2 is transferred as NEE from the Community Land Model Version 3 (CLM3)—coupled with the CASA′ terrestrial biogeochemistry model—to the Community Atmosphere Model (CAM3), where it is advected as a tracer throughout the troposphere. The underlying simulation is one of a number of transient runs performed for Phase 1 of the Coupled Climate/Carbon Cycle Model Intercomparison Project (C4MIP) using the Community Climate System Model (CCSM) on the Leadership Computing Facility (LCF) Computational Climate Science End Station (CCSES; Dr. Warren Washington, Principal Investigator). These CCSM simulations were run on the Cray X1E in the National Center for Computational Sciences (NCCS) at ORNL. Information about this study was provided by Dr. Inez Fung (LBNL), Dr. Jasmin John (University of California–Berkeley), and Dr. Forrest Hoffman (ORNL).
|
|
Summary and Future Directions
The questions being asked of the modeling community may be shifting from “Is climate change occurring?” to more quantitative “When?” and “How?” inquiries. To accurately address these questions and provide the best scientific groundwork for policy makers, many of the known deficiencies with coupled climate models must be fixed.
|
| Variations in the path of the Gulf Stream and other current systems need to be tracked to better predict heat transport, deep water formations, and melting under sea ice and ice sheets. Indeed, the CCSM needs to add a component for the world’s ice sheets before studies of rising sea level and rates of glacial accumulation and calving are theoretically grounded. In the atmosphere, precipitation must be brought more into line with geographic distributions so that soil moisture levels will be realistic and the right plants can grow in the right places. The coupling of ocean and atmosphere that give rise to oscillations like El Niño and the Pacific Decadal Oscillation (PDO) must also be improved. There is also hope that by including carbon and aerosol chemistry in the coupled model, the carbon cycle can be better quantified, understood, and predicted. |
| The next generation of models will earn the title “Earth system models.” This is, in fact, a major charge of the new SciDAC-2 project “A Scalable and Extensible Earth System Model for Climate Change Science,” led by PI Dr. John Drake (see “SciDAC-2: The Next Phase of Discovery,” p16). Supported by two Science Application Partnerships (sidebar “Science Application Partnerships,” p21), this new project aims to construct a first-generation Earth system model that fully simulates the coupling between the physical, chemical, and biogeochemical processes in the climate system. The broad web of investigators will build upon CCSM3 and looks forward to developing CCSM4 in time for the arrival of petascale machines planned for 2009. |
| The complexity of a new model, the consequent computer code, and the petascale simulation platform all present a great challenge to the scientists involved. Aside from the issues of scalability, researchers are also faced with a scarcity of data to validate biological responses and interactions within the climate system. The first Earth system model may not contain all possible and desired processes, but it marks an important beginning—the building of a solid base that will lead to scientific discovery. |
Contributors:
Dr. John Drake, ORNL; Dr. Philip Jones, LANL |
Further Reading
CCSM http://www.ccsm.ucar.edu/
|
