DYAMOND – Next Generation Climate Models

Nine models participated in the project, each simulating forty days and forty nights of the atmosphere’s and land-surfaces evolution. Daily sea surface temperatures and global meteorological analyses by the European Centre for Medium-Range Weather Forecasts (ECMWF) were provided as initial and boundary data. The simulated period, beginning on 1 August 2016 at 00 UTC, using global models with a storm-resolving grid spacing of 5 km or less, was chosen in coincidence with the NARVAL2 airborne field campaign that was initiated and led by MPI-M. For many models this was the first time they were ever run for such a long period and for such a realistic test. The comparison allowed an evaluation of their ability to robustly represent features of the climate system. This includes the representation of convective storms and clouds, and their response to solar forcing in the form of their diurnal cycle or the location of the Intertropical Convergence Zone (ITCZ). In addition to demonstrating an ability to better simulate such global storm-resolving runs over a longer period, the simulations provided output on scales that are much more comparable to what is observed. This apparent realism is shown in figure 1, where it is difficult to distinguish the satellite image from the ten models simulating the same time period.

The simulations performed as part of DYAMOND have additionally been used to outline the technical requirements necessary to bring this type of simulation capability to bear on the evolution of the climate system over decades, and as coupled to the ocean circulation. The DYAMOND experimental protocol was kept as simple as possible, with as few specifications for the participants as possible, to encourage participation and ensure a fast turn-around. The implementation took place within a year, which is exceptional for such model intercomparisons. Nine different groups from six national entities across three continents submitted results and demonstrated that such simulations are doable nowadays.

Before DYAMOND, most of the global storm-resolving models used here had never been run in this configuration, and no two models had ever before attempted to perform the same simulation. Preliminary analysis showed that basic aspects of the general circulation are well captured by the storm-resolving models. The outgoing long-wave radiation is consistent with observations, as is the global precipitation and distribution of precipitable water. Agreement among the models is such that it is difficult to establish if differences with observations—for instance, more precipitation in the poleward flank of the storm tracks and in the equatorial region just south of the Equator—represent deficiencies in the models or in the retrievals of these quantities.

Insofar as it can be determined given the brevity of the simulations and the basic storm-tracking algorithm, the tropical genesis and intensification statistics of tropical cyclones are quantitatively and qualitatively well captured. In addition, a diurnal cycle in global precipitation was identified that is consistent across all simulations and similarly evident in the observations, whereas it is inconsistent in climate models with a lower resolution. Considering that the models were not specifically tuned for this case and that minor implementation errors are to be expected given the early stage of the development, this agreement is remarkable.

The simulations are not without their deficiencies. Even the present, relatively superficial, analysis makes clear that there are still some challenges in trying to reap the rewards of being able to simulate the general circulation of Earth’s global atmosphere at storm-resolving scales.

One of the major challenges are shallow convection and stratocumulus clouds, which cannot yet be resolved by GSRM. However, they have a strong influence on the energy budget and sea surface temperatures. By virtue of the ability of GSRMs to directly link these processes to the circulation on the one hand, and observations at similar scales on the other, the scientists are optimistic that research targeted at addressing these challenges will bear fruit. If so, GSRMs will provide an improved physical foundation for model-based investigations of Earth’s climate and climate change and usher in a new generation of climate models.

Another important motivation for the DYAMOND project was to address performance and analysis bottlenecks associated with global storm-resolving models. The experiences across a variety of computer architectures with a variety of models suggest that a throughput of 6 to 10 days per day is feasible on a small 10 to 20 % of present day high-performance computers. GSRMs with up to 3 km global grids can be expected to reach a throughput of 100 simulated days per day in the coming few years. This rate of throughput makes such models applicable to a wider range of climate questions. Post-processing of the massive amounts of output from such models is challenging, both by virtue of its sheer volume and because of the more complex grid structures used by many models. In dealing with this output, domain-specific analysis tools, like CDOs (Climate Data Operators), proved to be invaluable and were optimized for the simulations.