Developing a low-resolution version of UKESM1

Colin Jones, Till Kuhlbrodt and Robin Smith, National Centre for Atmospheric Science (NCAS) and UKESM core group.

Alistair Sellar, Met Office Hadley Centre and UKESM core group.

Two versions of the UKESM1 model are under development, referred to as UKESM1-hr and UKESM1-lr. The planned resolution of these models is: UKESM1-hr; N216 (~60km) in the atmosphere and 0.25° in the ocean, UKESM1-lr; N96 (~140km) in the atmosphere and 1° in the ocean. Both models have 85 vertical levels in the atmosphere and 75 levels in the ocean. Due to the high resolution, UKESM1-hr is computationally expensive and therefore difficult to apply for long (centennial or greater) simulations when run with full Earth system process complexity. For similar reasons, large ensemble simulations using UKESM1-hr are also a challenge. Due to its coarser resolution, UKESM1-lr is ~6-10 times faster than UKESM1-hr and should help address these problems.

Figure 1Figure 1. Zonal mean Transient Eddy Kinetic Energy (DJF) in N48 (top left), N48 minus N96 (top right), N96 minus MERRA reanalysis (bottom left) and N48 minus MERRA reanalysis (bottom right).

This article discusses the formulation of UKESM1-lr, in particular work carried out over the past 12 months to determine a suitable model resolution. This work is detailed in the report: Recommending a resolution for UKESM-LO (PDF, 6MB) (Kuhlbrodt et al. 2015), recently circulated to interested Earth system modellers in the UK. Here we provide a brief overview of the main findings of the report and outline how the model will be used in the coming years, particularly in the context of CMIP6.

The resolution choice for UKESM1-lr is primarily a compromise between: (i) computational speed, (ii) sufficient model resolution and (iii) sufficient parameterisation complexity. With respect to option (iii), for a given application where both models are used we aim to maintain scientific traceability between them. To the degree possible parameterisation complexity should therefore be common, except when differences in resolution demand a specific parametric representation of an unresolved process. An example of this is the need to include the Gent and McWilliams (1990) parameterization of the transport by oceanic mesoscale eddies that are “permitted” at 0.25° resolution but definitely absent (unresolved) at 1° resolution.

This leaves computational speed and sufficient model resolution as the two key considerations. These have been assessed starting from the coupled physical model HadGEM3 (GC2: Global Configuration 2). Resolution choices largely concerned N48 (~280km) or N96 (~140km) for the Unified Model (UM) atmosphere and 1° or 2° for the ocean model NEMO. The UM at N48 is significantly faster than N96 resolution. Nevertheless, extensive evaluation of N48 simulations clearly shows this model to be significantly worse, in terms of science quality, than N96. In particular, mid-latitude storm track dynamics (represented by upper troposphere transient eddy kinetic energy) are significantly worse at N48 (figure 1) as are the majority of standard model performance metrics (figure 2). More detail on the comparison between N48 and N96 can be found in Kuhlbrodt et al. (2015). Initial tests of the N96/1° coupled configuration show the UM atmosphere constitutes the bulk of the computational cost of this model version, hence when using an N96 UM atmosphere there is little computational benefit from choosing NEMO 2° over 1°. Given the known improvement in simulated ocean dynamics at 1° relative to 2° we decided the ocean resolution of UKESM1-lr should be 1°, resulting in a recommended resolution of N96/1° for the physical core of UKESM1-lr.

Figure 2Figure 2. Performance metrics comparing N48 (coloured dots) against an N96 reference. N48 metrics are referenced to a value of unity representing N96 performance. Red dots indicate where N48 is significantly worse than N96. Orange dots show where differences are not significant and green dots where the performance of N48 is within observational uncertainty. Top left: Global Troposphere metrics, Bottom left: Global Radiation metrics, Top Right: Mid-latitude storm track metrics, Bottom Left: Mid latitude blocking metrics.

A further reason for choosing UM N96 over UM N48 is that the N48 configuration is no longer an officially supported version at the Met Office. This led to significant technical challenges in generating new ancillary files for the N48 model version and was one reason we began all analysis of N48 in atmosphere-only mode (since creating a coupled version of UM N48 and NEMO would have been a major technical effort). Considering UKESM1-lr will be used over the approximate period 2016-2022, including in a large number of CMIP6 Model Intercomparison Projects (MIPs), we concluded the lack of support for a N48 UM configuration would be an enormous drain on both the UKESM core group and UK scientists wishing to perform science experiments with the model.

Subsequent to this analysis a significant effort is now under way to try and increase the computational throughput of the N96 UM. Tests have been made to lengthen the model dynamical time step from the standard 20 minutes, with successful tests at 30 and 45 minutes (the latter proving difficult to use in practice due to the assumed output frequency in the UM STASH diagnostics) and, preliminary but not yet successful, tests of a 60 minute time step. We have also been assessing the computational benefit resulting from increasing the UM radiation time step from the present 1 hour to 3 hour, with a 1-hour incremental update of the incoming solar radiation flux (see Manners et al. 2009). These changes indicate the potential to increase the overall throughput of the physical core of UKESM1-lr at N96/1° from an initial 4.5 simulated years per day to ~10-12 years/day. Work is continuing to assess the overall science changes arising from these longer dynamical and radiation time steps, before any final decision on time step length is made. This analysis may need to be repeated once UKESM1-lr Earth system component models are coupled to the coupled physical model. Once such an assessment is complete, UKESM1-lr will be released with a recommended set of model time steps. Similar analysis will also be carried out with respect to the most efficient processor distribution between the atmosphere and ocean model executables and a set of recommended processor configurations will be included with the model release.

We aim to have UKESM1-lr scientifically ready for use in CMIP6 by August/September 2016. At this point we will begin a set of mandatory CMIP-DECK simulations (Meehl et al. 2014), including a long (~500 year) pre-industrial control and an historical (1850-present day) run. Parallel to these integrations we will release UKESM1-lr to the UK community and begin using this model in a number of CMIP6 MIP experiments. A key aspect of the overall evaluation of UKESM1-lr will be the level of performance traceability with UKESM1-hr. This analysis has begun and some initial results are reported in Kuhlbrodt et al. (2015). More evaluation and comparison of both UKESM1-lr and -hr will continue over the coming 12-24 months.

REFERENCES
  • Kuhlbrodt, T., C. Jones, A. Sellar and R. S. Smith (2015), Recommending a resolution for UKESM-LO. A UKESM project report. Met Office and NERC.
  • Manners, J., Thelen, J.-C., Petch, J., Hill, P. and Edwards, J.M. (2009), Two fast radiative transfer methods to improve the temporal sampling of clouds in numerical weather prediction and climate models. Q.J.R. Meteorol. Soc., 135: 457-468. doi: 10.1002/qj.385
  • Meehl, G. A., R. Moss, K. E. Taylor, V. Eyring, R. J. Stouffer, S. Bony, and B. Stevens, Climate Model Intercomparison: Preparing for the Next Phase, Eos, Trans. AGU, 95(9), 77, 2014.

Download this article as a PDF