General circulation models (GCMs) are the main computational tools for theoretical research on global climate variability and climate prediction. Such models are the intellectual direct descendants of the numerical weather prediction models envisioned early in this century by Lewis F. Richardson in his book, Weather Prediction by Numerical Process (Cambridge University Press, 1922). We may distinguish between atmospheric GCMs (AGCMs), oceanic GCMs (OGCMs), and coupled GCMs, which combine these two major components of the climate system. The clear direction of the evolution of these models is toward coupled, comprehensive models of the entire Earth system, including biogeochemical aspects. At present, AGCMs are the most advanced models, having now undergone four decades of continuous development. Much worthwhile climate research can be performed with AGCMs, either in stand-alone mode, driven by fixed sea surface temperature (SST) boundary conditions, or coupled to any of several types of ocean models which allow SST to vary.
The number of available AGCMs has increased rapidly in recent years. In the early 1960s, only 3 state-of-the-art models existed worldwide, and all were in the U.S. By the mid-1980s, that number had increased by at least an order of magnitude, as supercomputer resources became commonplace in many countries. Now the number of GCMs and of scientists involved in GCM research is growing explosively, driven both by the growth in computing power at the workstation level and by the public dissemination of GCM codes. These codes, which were typically treated as proprietary by their creators for many years, are now becoming widely available. For example, the AGCM recently developed in the United States at the National Center for Atmospheric Research (NCAR), known as the Community Climate Model, Version 2 (CCM2), was first ported to IBM RISC Unix workstations from its original (Cray supercomputer) development platform by NCAR staff, and then the source code for both versions was made publicly available in 1994. This code is now easily and instantly obtainable by anyone with Internet access. Thus, a 50,000 line (Fortran code) modern AGCM, representing many person-years of development effort, professionally programmed and fully documented, is now available at no cost to the entire global change research community.
As one typical example of the resulting increase in the population of GCM users, CCM2 is now running on Digital Equipment Corporation workstations at the Climate Research Division of Scripps Institution of Oceanography, University of California, San Diego. The conversion from the IBM version of the code was carried out at Scripps in 6 weeks by a postdoctoral fellow. Thus, GCM work, which formerly required not only supercomputer access, but also relatively large scientific staffs and specialized computational expertise, has now become possible for typical academic researchers, working in the mode of a single professor collaborating with one or two graduate students or postdoctoral fellows. This democratization of GCM research is now giving rise to qualitative changes in the way the research is conducted. For example, email networks are springing up, connecting CCM2 users, who trade reports of code bugs, revised physical process parameterizations, and the like. At the same time, GCM research is becoming less conservative intellectually, as the reduced cost motivates scientists to develop and test unconventional parameterizations, modified numerical algorithms, and other departures from past practice. These trends are likely to continue and accelerate, as more codes become public and as the cost of computing continues to decline.
A second driver which is changing the way climate modelers work is the proliferation of observational data sets for the development and validation of GCMs. In the early days of GCM research, models were compared against only a few sparse data sets, typically time-averaged and in two space dimensions, representing the globally mapped or zonally averaged fields of the primary meteorological variables, such as sea-level pressure, geopotential height of standard pressure surfaces, temperatures, humidities, and winds. Today, available data sets cover a far greater array of field variables, and they span three space dimensions and time. Disparate data sources, ranging from paleoclimate data to analyzed fields from operational numerical weather prediction to specialized satellite remote sensing measurements, are being combined in heterogeneous data sets and used routinely in GCM research. Again, the clear trend is toward even larger and more comprehensive data sets, as the sources of data continue to increase. Similarly, the data produced by the GCMs themselves is also growing rapidly in size and variety, as model resolution and physical complexity both increase.
These two factors, the proliferation of GCMs and the increase in the size and scope of the associated data sets, mandate a change in the way GCM researchers work. Antiquated ways of handling model and observational data are hamstringing researchers now, and the seriousness of the problem is certain to increase in the future. Nevertheless, the new modes of GCM research are already beginning to bear fruit in the form of new scientific insights. At Scripps, Somerville and Wan-Ho Lee have recently used an AGCM (created by using their own parameterizations to replace those of CCM2) to test a suite of alternative cloud-radiation algorithms. As is well known, most of the differences in GCM sensitivity to increased greenhouse gas concentrations, when measured by global average equilibrium surface temperature changes, are due to the differing cloud and radiation algorithms in use by the various modeling groups. The work of Somerville and Wan-Ho Lee extends and generalizes the study of Senior and Mitchell (J. Climate, 6:393-418, 1993) who used a version of the AGCM developed at the United Kingdom Meteorological Office.
The methodology of Somerville and colleagues relies on using both a coupled model (the AGCM plus their own simple ocean mixed layer model) and a set of perpetual July AGCM integrations forced by constant SST perturbations, following Cess et al. (J. Geophys. Res., 95:16,601-16,615, 1990). They have tested parameterizations including relative humidity-based clouds and several versions of schemes involving a prognostic cloud water budget. They have carried out extensive sensitivity tests with these parameterizations, in which they examine the effect of varying tunable constants and other arbitrary aspects of the schemes. They compared the GCM results with Earth Radiation Budget Experiment (ERBE) data and with observations from the U. S. Department of Energy's Atmospheric Radiation Measurement (ARM) Program, diagnosed using the single-column models developed by Iacobellis and Somerville (J. Atmos. Sci ., 48:1948-1959 and 1960-1971, 1991). In general, the parameterizations with computed radiative properties based on cloud water budgets are better able to reproduce ERBE observations of cloud forcing. They also tend to give qualitatively different cloud feed backs (for both solar and terrestrial radiation) when compared to parameterizations based on relative humidity clouds with fixed radiative properties.
Plans for future research on the development, validation and improvement of numerical climate models will emphasize the use of GCMs together with diagnostic models and field program observational data for the direct validation of physical process parameterizations. The research focuses on the parameterization of atmospheric convection and cloud-radiation interactions and has three major components: