2.1 The UKA3 atmosphere component

The atmosphere model component within UKC3 (named UKA3 when run in atmosphere-only mode) uses the RA1-M regional atmosphere science configuration described in detail by Bush et al. (2019). This is implemented using the MetUM code at version 10.6 (e.g. Walters et al., 2017). Table 4 highlights the key similarities and differences between UKA3 and UKA2 configurations. Updating between MetUM model code vn10.1 and vn10.6 introduces a substantial number of incremental scientific and technical fixes, enhancements and optimizations, delivered through ongoing model evaluation and development. Technical details on the RA1 science configuration are also provided to registered users of the MetUM code at https://code.metoffice.gov.uk/trac/rmed/wiki/ra1 (last access: 18 February 2019). Key changes introduced between UKA3 (RA1) and the UKA2 (PS37) science configuration in model code or namelist options used of relevance to the regional atmosphere performance are highlighted below. The corresponding namelist changes are listed in Table 5.

Table 5Summary of key changes between UKA2 and UKA3 configuration MetUM namelists, associated with implementing enhanced science options based on regional atmosphere-only model development and evaluation. Registered MetUM users can access further details at https://code.metoffice.gov.uk/trac/rmed/wiki/ra1/protoRA1 (last access: 18 February 2019).

Download Print Version | Download XLSX

Modifications to MetUM for regional atmosphere coupling

As for UKC2, exactly the same codes are compiled and built for both coupled and uncoupled configurations to ensure all simulations are run with identically built code. A number of required code adaptations were implemented as branches to the vn10.6 MetUM trunk code. These are detailed for reference in the Supplement. In general, these modifications can be categorized as being required either to

• apply the RA1-M graupel definition in the JULES snow scheme (Sect. 2.2),

• couple effectively between ocean and atmosphere grids with mismatching coastlines, due to grid interpolation of mismatched land/sea masks (as in Lewis et al., 2018),

• enable dynamic coupling and exchange of information between the atmosphere and a wave model,

• enable river routing within the coupled MetUM–JULES system,

• enable consistent coupling of snow when convective snow is explicitly resolved.

To encourage collaboration, a single merged copy of the UKA3 MetUM model code is available to registered researchers via a shared code repository, which can be accessed via https://code.metoffice.gov.uk/trac/utils/browser/ukeputils/trunk/gmd-2018 (last access: 18 February 2019). The code repository location is also linked directly for registered researchers in Table 4.

Table 6Summary of UKA3 land surface component, and key similarities and differences to the UKA2 configuration described by Lewis et al. (2018). The direct links to merged code are provided to support collaboration with registered researchers. Further information on accessing JULES can be found at http://jules.jchmr.org (last access: 18 February 2019).

Download Print Version | Download XLSX

2.2 The UKA3 land surface component

The JULES land surface component within UKC3 is implicitly coupled to the MetUM atmosphere model code, using the method of Best et al. (2004), in all configurations with an atmosphere component. This combined atmosphere–land configuration is termed UKA3. Table 6 lists the key similarities and differences between the land surface specification in UKA3 and UKA2. Details of the trunk code updates between JULES releases vn4.2 and vn4.7 can be accessed at http://jules-lsm.github.io/vn4.7/ (last access: 18 February 2019). The majority of changes, however, are not considered relevant for the regional land surface component. The UKA3 land surface definition is a direct implementation of the RA1-M science configuration, with additional options and parameters enabled for river routing. Key changes introduced in UKA3 are highlighted below, with corresponding namelist changes given in Table 7.

Modifications to JULES for regional coupling

Similar to the UKC2 system, a number of required code adaptations were implemented as branches to the vn4.7 JULES trunk code to enable regional coupling and run the MetUM–JULES coupled system with river routing. These are detailed for reference in the Supplement. In general, these modifications can be categorized as being required either to

• apply the RA1 graupel definition in the JULES snow scheme,

• enable dynamic coupling and exchange of information between the atmosphere and a wave model,

• apply the Martínez-de la Torre et al. (2018) approach of a slope-dependent Probability Distribution Model runoff generation,

• enable river routing within the coupled MetUM–JULES system,

• apply a check in the calculation of surface exchange coefficients for slightly unstable conditions.

A single merged copy of the UKA3 JULES model code is also made available to registered researchers via a shared code repository, which can be accessed via https://code.metoffice.gov.uk/trac/utils/browser/ukeputils/trunk/gmd-2018 (last access: 18 February 2019). The repository location is also linked in Table 7.

Table 7Summary of key changes between UKA2 and UKA3 configuration JULES namelists, associated with implementing enhanced science options based on regional atmosphere-only model development and evaluation. Registered MetUM users can access further details at https://code.metoffice.gov.uk/trac/rmed/wiki/ra1/protoRA1 (last access: 18 February 2019).

Download Print Version | Download XLSX

2.3 The UKO3 ocean component

The most significant change related to coupling introduced between UKC2 and UKC3 is the implementation and configuration of wave-to-ocean feedbacks within the NEMO ocean model code. Further details are provided in Sect. 3.

Table 8Summary of UKO3 ocean component, and key similarities and differences to the UKO2 configuration described by Lewis et al. (2018). Note that the NEMO compilation key key_harm_ana was only used in UKO3 implementations.

^* See the Supplement for description of compilation keys. The direct links to merged code are provided to support collaboration with registered researchers. Further information on accessing NEMO can be accessed at http://www.nemo-ocean.eu (last access: 18 February 2019).

Download Print Version | Download XLSX

Table 8 highlights other common and differing aspects of the UKO3 regional shelf-seas ocean-only configuration relative to UKO2. Updates were introduced in order to maintain a common science configuration to the evolving Atlantic Margin Model (AMM15) ocean-only shelf-seas forecasting system, which is described in detail by Graham et al. (2018). This required updating the NEMO vn3.6 trunk code revision from r5518 to r6232, which includes a number of minor bug fixes and technical code improvements only. The following configuration changes were also implemented.

Updated Baltic Sea boundary condition

For simplicity in UKO2, the inflow to the domain from the Baltic Sea at the eastern boundary was treated as two river sources, located in the Kattegat Strait. In UKO3, as in AMM15, eastern boundary conditions are instead taken from a regional Baltic simulation (Gräwe et al., 2015). Baltic boundary conditions are applied over a relaxation zone of horizontal width (nn_rimwidth) of 10 grid cells, while boundary conditions into the majority of the domain along the remaining edges are applied over a relaxation zone of 15 grid cells.

Table 9Summary of UKW3 WAVEWATCH III wave model component, highlighting substantive aspects same as for UKW2 configuration described by Lewis et al. (2018).

^* A fuller description of the compilation switched is provided in the Supplement. The direct links to merged code are provided to support collaboration with registered researchers. Further information on accessing WAVEWATCH III can be found at http://polar.ncep.noaa.gov/waves/wavewatch (last access: 25 February 2019).

Download Print Version | Download XLSX

2.4 The UKW3 surface wave component

Table 9 reflects the close similarity in configuration between UKW2 and UKW3 wave model components. This is a result of model code developments being limited to the provision of new coupled fields to support wave-to-ocean coupling (Sect. 3) and some minor bug fixes. A copy of the WAVEWATCH III model code used to define the UKW3 configuration is made available via https://code.metoffice.gov.uk/trac/utils/browser/ukeputils/trunk/gmd-2018 (last access: 18 February 2019) to researchers who are registered as WAVEWATCH III users. The exact path is also linked from Table 9.

3.1 Momentum modified by drag coefficient (atmospheric forcing modes only)

The wind stress from the atmosphere at the ocean surface that is transmitted to the ocean is modified due to the wave roughness. In a fully coupled system in which the atmospheric boundary layer is modified by the wave state (e.g. UKC3aow), it is considered that this effect is simulated through the wave–atmosphere coupling, and so the ocean component is driven by a wave-modified atmosphere. However, for partially coupled (e.g. UKC3ow) or wave-forced (e.g. UKO3gw) ocean configurations, it is possible to account for the wave-modified drag in computing the wind stress acting on the ocean.

In the UKO3 shelf-sea configuration (with ln_shelf_flx = .true. in namelist namsbc_flx), the wind components are read, and the wind stress is typically calculated from them using a drag coefficient which is a function of the wind velocity (nn_drag = 0) according to Eq. (1; Smith and Banke, 1975). In the new coupled or wave-forced implementation (ln_cdgw = .true.), the shelf sea configuration is also used and the drag coefficient C_D is calculated by the wave model (nn_drag = 1), and applied as shown in Eq. (2). This formulation can also be used (nn_drag = 2) with a constant value for C_D, set at 0.0015.

Alternative implementations of this effect are also available for use with bulk forcing mode configurations of NEMO.

As highlighted by Eq. (1), C_D is a function of wind speed. Figure 2b and e illustrate the distribution of C_D computed by the UKW3g WAVEWATCH III wave model configuration for a summer and winter month, and Fig. 2h shows its simulated variation as a function of forcing wind speed at a selected point in the North Sea. This shows values of between approximately 0.005 and 0.0025. The nn_drag = 2 default constant of 0.0015 appears to correspond to forcing winds of approximately 10 m s⁻¹. There is a general tendency for the wave-simulated C_D to exceed the default nn_drag = 0 formulation (i.e. Eq. 1; Smith and Banke, 1975) for wind speeds in excess of approximately 5 m s⁻¹, and produce lower values for slower wind speeds. Results from UKC3ow and UKW3go simulations (not shown) also confirm generally low sensitivity of C_D values to the presence of ocean forcing or coupling in the wave model.

3.2 Momentum fraction transferred to the ocean through wave breaking

Part of the momentum that the ocean receives from the atmosphere (after taking into account the effect of wave roughness either through wave–atmosphere coupling or modifying the drag coefficient through Sect. 3.1) is stored in surface waves through wave growth or released from the surface waves on wave breaking. The momentum that will actually force the ocean is therefore a fraction of the atmospheric momentum, which is calculated within the wave model according to Eq. (3). The fraction of atmospheric momentum transferred to the ocean is approximated by the normalized momentum flux variable tauoc (Breivik et al., 2015).

The following definitions are used to describe each component of the surface momentum budget.

• τ_atm: stress applied by atmosphere on ocean surface

• τ_wav: momentum flux absorbed by wave field

• τ_wav:ocn: momentum stored by waves released to ocean through wave breaking

• τ_ocn: water-side stress transmitted into the ocean

Note that, as in Breivik et al. (2015), τ_wav and τ_wav:ocn terms are computed from the wave model source terms across the model's frequency range only. This implies that input and dissipation are balanced at higher frequencies. Further work is required to fully account for the tails of the wave frequency range.

Figure 2a and d show the mean simulated tauoc for a summer and winter month, and highlighted values tend to lie in the range 0.95 to 1.05 (i.e. approximately 5 % modification to the atmosphere surface stress due to waves). Largest enhancement can be found along west-facing coastlines, and largest reductions in the lee of land such as downstream of the Scottish islands, in the Irish Sea and along the English Channel. The spatial distribution is broadly consistent between summer and winter months, but with the magnitude of wave modification clearly increased in winter.

3.3 Stokes–Coriolis drift

The Stokes drift, caused by finite amplitude waves, creates a relative motion along the wave direction which quickly decays with depth. The NEMO momentum equation is modified to account for the Stokes drift velocity v_s, taking into account the Coriolis forcing, as in Eq. (4).

As only the surface Stokes drift, v₀, is usually available from the wave model, different parameterizations are used to estimate the change in the Stokes drift velocity with depth, v_s(z), as a function of the mean wave period, t₀₁; significant wave height, H_s; and peak wave frequency ω_p. Options are controlled by the nn_sdrift NEMO namelist parameter.

For nn_sdrift = 0, the Breivik 2015 parameterization is used (Breivik et al., 2015, Eq. 120), with the Stokes drift velocity profile v_s(z) given by Eq. (5). If nn_sdrift = 1, the Phillips parameterization (Breivik et al., 2016, Eq. 100) is applied using an inverse depth scale, according to Eq. (6). An extension can be applied if nn_sdrift = 2 using the peak wave number as calculated by the wave model rather than the inverse depth scale, as shown in Eq. (7).

Figure 2c, f and i also show a summer and winter spatial distribution of the surface Stokes drift velocity and its variability with wind speed.

3.4 Wave-modified surface roughness

The ocean surface roughness, which has an effect in the vertical mixing by defining the surface turbulent mixing length scale, can be calculated in different ways. In the generic length scale (GLS) turbulent closure scheme, used in the UKO3 configurations; this is dependent on the choice of parameter nn_z0_met. For example, Eqs. (8) and (9) show the simplest approach defining either a constant roughness or constant Charnock parameter via namelist settings.

Rascle et al. (2008) discuss how the roughness length is more physically related to the scale of breaking waves and related eddies responsible for high mixing levels close to the surface, and that it can be related to the significant wave height H_s (or more strictly the wind-sea wave height).

By default in UKO3 ocean-only configurations, the Rascle et al. (2008) parameterization for H_s (nn_z0_met = 2) is used. This estimates the wave age $C_{\text{p}}/u_{\ast }$ and subsequently significant wave height H_s as a function of wind speed, through Eq. (10a) and (10b).

Here $u_{{\ast }_{\text{ref}}}$ is a typical friction velocity (0.3 m s⁻¹). The surface roughness length z₀ is then taken as a fraction rn_frac_hs (default value 1.3) of the estimated H_s. The appropriate value for this factor for the north-west shelf region should be reviewed further before consideration of this scheme for operational applications, while further development might consider the coupling of the wind-sea wave height computed by WAVEWATCH III explicitly.

When using wave forcing or coupling (nn_z0_met = 3), the same approach is used but with the wave model significant wave height replacing the Rascle et al. (2008) estimate, according to Eq. (11).

4.1 Evaluation framework

Lewis et al. (2018) described the development of an evaluation framework to understand the performance of model components run within coupled systems relative to uncoupled approaches. All coupled and uncoupled configurations defined in Table 2 are provided as Rose suites (http://metomi.github.io/rose/doc/rose.html, last access: 18 February 2019) and version controlled under the Flexible Configuration Management (FCM) system (http://metomi.github.io/fcm/doc/, last access: 18 February 2019). A number of different options for initial conditions or forcing are available for each configuration (Table 2), which are enabled within a given suite by setting the relevant RUNID environment variable.

4.2 Model experiments

The focus of UKC3 system evaluation discussed in this paper is on a series of four coupled and forced simulation experiments, each of approximately 1-month duration. This enables assessment across a variety of meteorological conditions within a given month and of ocean and wave results across a spring–neap tidal cycle, and covers evaluation at different times of the year. In order to capture a range of conditions, experiments are conducted for the following periods:

• a.

  “Spring”: 30 March–19 April 2014

• b.

  “Summer”: 30 June–30 July 2014

• c.

  “Autumn”: 30 September–30 October 2014

• d.

  “Winter”: 30 January–28 February 2015

This approach extends the analysis of Lewis et al. (2018) who considered a number of relatively short 5-day case study simulations across a range of conditions. By extending the simulation period, it is hoped to provide a more comprehensive evaluation of system performance and to establish whether any long-term drifts develop in any component or have an impact on the coupled system overall. These experiments represent the first time for the UK regional coupled prediction system to be run for such duration, and provides a useful check on the feasibility of applying the UKC3 or its successor systems for longer-term applications such as generation of climate scenarios. In general, it is found that the coupled system remains stable over the month-long simulation period across ocean, wave and atmosphere components, with no serious model drifts found, even without data assimilation. This gives confidence to the scientific validity of the configurations developed and suggests these to be suitable tools for conducting even longer-duration research runs. This stability can be partly attributed to the forcing of each component at the lateral boundaries of the regional domain, and the use of a fixed climatology for river flows into the ocean.

Table 10Summary of model configurations used for system evaluation simulations relevant to each model component. See also Table 2 for configuration definitions.

Download Print Version | Download XLSX

Table 10 lists the model simulations conducted for each period. Given the increased computational cost of running model experiments over longer periods, only a subset of the possible system configurations defined in Table 2 are considered here. Coupled model results from fully (UKC3aow) and partially (UKC3ao, UKC3owg) coupled mode simulations are compared with the UKA3u, UKO3g and UKW3g control simulations, as these are considered to be the most analogous to typical operational configurations currently in use.

It is considered most efficient to focus results on a relative evaluation between coupled and uncoupled simulations of the UKC3 system over the selected periods, rather than on a comparison between the UKC2 and UKC3 releases. This approach helps to isolate the impact of the new coupling capabilities within UKC3 from any other code or configuration updates between versions, and provides a more relevant summary of the relative performance of the coupled system relative to current operational approaches using more recent science configurations and code.

All simulations are initialized with the same initial conditions relevant to each model component, and the lateral boundary conditions applied are common across simulations, irrespective of the mode of running.

The following analysis compares model outputs to a variety of in situ observations taken from the Met Office operational archive. Figure 1 illustrates the typical data availability, as available for assessment of 1 day of the “Summer” 2014 experiment. Further details on the observing networks are provided by Lewis et al. (2018). Where model results are compared with observations, a “neighbourhood” of 3 by 3 grid cells nearest the observation location are considered and local mean values computed. This is considered to provide a more robust evaluation than considering only the nearest matching model grid cell for kilometre-scale model systems (e.g. Mittermaier, 2014).

Figure 3Sensitivity of ocean model sea surface temperature (SST) to coupling. (a, d, g, j) Monthly mean difference between fully coupled and ocean only (UKC3aow–UKO3g) during April, July, October 2014 and February 2015 runs, respectively. (b, e, h, k) Monthly mean difference between fully coupled and partially coupled (UKC3aow–UKC3ao) runs during each experiment. Note the different colour scale. (c, f, i, l) Percentage difference in surface temperature RMSE statistic for UKC3aow relative to UKO3g.

Download

4.3 Ocean component results

Figure 3 shows a summary of experiment-mean ocean model SST results across each simulation period, comparing fully coupled UKC3aow, partially coupled UKC3ao and forced-mode UKO3g configurations. The sensitivity of SST to coupling is highest during summer months as expected (e.g. Lewis et al., 2018).

The mean differences UKC3aow–UKO3g represent the impact of full atmosphere–ocean–wave coupling relative to a free-running ocean-only configuration. To a first order approximation, differences are therefore a combination of the impact of wave forcing and feedbacks on the ocean, the impact of a change in meteorological forcing resulting from increased atmospheric resolution from global (∼17 km) to regional (1.5 km) scale, and the effect of three-way coupled feedbacks between ocean and atmosphere, ocean or waves. In April, July and October runs, the impact of full coupling is a mean reduction of SST. This is typically 0.2 K in April and October, but mean changes of up to 1 K are found for the July 2014 simulations (Fig. 3d). The relative impact of wave feedbacks on both the ocean and atmosphere is illustrated in Fig. 3b, e, h and k by comparing UKC3aow with UKC3ao. This highlights a general tendency for wave coupling to cool the simulated SST, by up to 0.5 K, which is found to be mostly driven by a relatively reduced surface drag in April, July and October at least (e.g. Fig. 2b). This effect can be replicated in the partially coupled UKC3owg configuration through the wave-modified surface drag (Sect. 3.1).

The comparison between model results and in situ SST observations presented in Fig. 3 shows a general improvement in RMSE statistics, particularly at the near-coastal buoys but also more widely for UKC3aow relative to UKO3g. Time series of the average ocean model bias for SST through each simulation are shown in Fig. 4. This highlights that the UKO3g ocean-only simulation is generally biased warm for each season, and by up to 1 K during the July 2014 run. Note that all simulations are initialized from a multi-annual run of the UKO3g ocean configuration. For the July 2014 experiment, while UKO3g results maintain the initial bias of approximately 1 K too warm throughout the month-long simulation, the initial bias is eroded over the first week or so of simulations to be approximately 0.2 K too warm in both UKC3ao and UKC3aow runs, and is further reduced in the last week of the simulation. Similar features but of smaller magnitude can be seen during April and October 2014 experiments. This result is found to be a function of both the spatial and/or temporal resolution change (noting that the UKO3g forcing is obtained from operational archives with data assimilation) in the atmosphere forcing, and a result of an improved atmospheric state due to the dynamic coupling to the ocean (and wave) component. The magnitude of improvements due to coupled relative to observations is further highlighted for these months in Fig. 4b, d and f which show the time series of the absolute model bias relative to the UKO3g control.

The contribution of wave processes to the reduction in SST bias can be determined in fully coupled mode by comparing UKC3aow and UKC3ao results (dark red relative to light red in Fig. 4), and in partially coupled mode with consistent atmospheric forcing by comparing UKC3owg with ocean-only UKO3g (blue relative to grey line in Fig. 4). This shows representation of wave processes to improve the agreement of simulated SST with observations, but also that this contribution is estimated to be approximately 10 % of the differences between fully coupled and ocean-only modes. A larger relative improvement is also found in forced mode (UKC3owg–UKO3g) than found between UKC3aow and UKC3ao in April and July 2014 at least.

Figure 4(a, c, e, g) Time series of average model–observation (MODEL–OBS) bias of ocean model sea surface temperature across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to UKA3u. A negative relative |average bias| indicates the coupled system to have a lower average absolute bias across all observation sites than the control. Note that plots have different scales across the different months evaluated, and observations are compared with a nearest neighbourhood mean of 3 by 3 model grid cells.

Download

Results for a winter month (February 2015) show the impact of full coupling and wave feedbacks to be more isolated to near-coastal areas, and while improvements in RMSE for UKC3aow relative to UKO3g can be seen in Fig. 3l, the relative difference in mean bias in Fig. 4h fluctuates through the month (but typically within 0.1 K). In this case, the impact of wave feedbacks is generally very similar between UKC3aow and UKC3owg simulations relative to UKO3g.

This discussion highlights the potential of regional coupled systems to deliver improved simulation of the ocean state, and this development is being tested for implementation within the framework of the EU Copernicus Marine Environment Monitoring Service (CMEMS) for the north-west European shelf region for example.

Figure 5Sensitivity of wave model wind forcing, significant wave height (H_s) and peak wave period (T_p) to coupling during October 2014 experiment. Monthly mean differences between UKC3aow and UKW3g for (a) coupled/forced atmosphere winds, (b) significant wave height H_s and (c) peak wave period T_p. (d–f) Percentage difference in RMSE statistic for UKC3aow results relative to UKW3g for (d) wind forcing, (e) H_s, (f) T_p. (g–i) Monthly mean differences between partially coupled UKC3owg and UKW3g (i.e. with same global-scale wind forcing) for (g) wind forcing, (h) H_s, (i) T_p, and (j–l) percentage difference in RMSE statistics for each variable for UKC3owg relative to UKW3g.

Download

4.4 Wave component results

It is widely known that the quality of the wave model results is critically dependent on the quality of the wind forcing, or that applied via atmosphere–wave coupling (e.g. Cavaleri et al., 2018). The benefit of coupling on wave results was therefore characterized by Lewis et al. (2018) for case study simulations through comparing UKC2aow results with the UKW2h control using a comparably high-resolution wind forcing, and it was found to be generally difficult to improve on the performance of the wave-only simulations forced by operational archive global resolution MetUM winds in the UKW2g configuration. The same characteristics are found for UKC3 results in this study, in which the fully coupled UKC3aow runs are compared only with partially coupled UKC3owg and forced mode UKW3g simulations in which the wind forcing is provided by the global resolution operational MetUM archive. Figure 5 illustrates the differences in wind forcing during the October 2014 experiment, which is representative of results for other months. The impact of atmosphere model resolution can be seen in particular around the coasts, where the UKC3aow winds exceed those from the global MetUM forcing. This is a combination of the effect of increased drag over land impacting a broader region in the global-scale forcing than at high resolution, and from including currents and tidal feedbacks in the lower boundary condition in the coupled configuration.

Figure 6Difference of wind forcing applied to wave model. (a, c, e, g) Time series of average model–observation (MODEL–OBS) bias across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to the UKW3g wind forcing. Note that plots have different scales across the different months evaluated, and observations are compared with a nearest neighbourhood mean of 3 by 3 model grid cells.

Download

Wind speeds tend to be slightly reduced in regions away from coastlines across the north-western and south-western approaches to the UK, and across the southern North Sea. These features are generally replicated during other months.

Comparison with available in situ wind observations located in the ocean (Fig. 5d), noting these are generally located away from near-coastal areas, demonstrate a consistent reduction in quality of winds at high resolution in UKC3aow relative to the global archive MetUM forcing. Figure 6 presents a more quantitative analysis through each experiment, and shows that, on average, the UKW3g (and UKC3owg) forcing is biased fast by up to 1 m s⁻¹ during each run across the domain. The fully coupled UKC3aow winds are by contrast biased fast by up to 2 m s⁻¹ during the four periods, but with much greater variability in the magnitude and sign of the bias through time than found for the UKW3g MetUM forcing. The fast bias is consistent with the recent analysis by Jiménez and Duidha (2018) who compared WRF model simulations with observations from the FINO (Forschungsplattformen in Nord- und Ostsee) towers located in the southern North Sea.

Figure 5b demonstrates the close link between wind speed differences between configurations and their impact on significant wave height, H_s, with increased wind speeds in UKC3aow over the northern North Sea on average tending to drive waves of increased magnitude, and reduced wave heights to the west of the UK in the fully coupled UKC3aow simulations relative to the wave-only UKW3g control. Figure 5e also highlights an expected general reduction in the level of agreement between UKC3aow wave height simulations with in situ observations relative to UKW3g. However, even given the degraded wind speed forcing, relative improvement can be seen at a number of near-coastal sites along the south-western English Channel coast.

To isolate the impact of wave–ocean feedbacks from the wind forcing, Fig. 5h and k compares UKC3owg with UKW3g results. This shows a general reduction of significant wave height in most areas, but a region of slightly enhanced wave heights along the northern half of the English Channel, associated with wave–current interactions. The impact of current–wave interactions in the near coastal zone was also discussed by Lewis et al. (2018). In contrast to Fig. 5e, wave–ocean interaction is shown to have a clear beneficial impact on the agreement between observed and simulated wave height in Fig. 5k.

Figure 7Sensitivity of wave model significant wave height (H_s) to coupling. (a, c, e, g) Time series of average model–observation (MODEL–OBS) bias across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to UKW3g. Note that plots have different scales across the different months evaluated, and observations are compared with a nearest neighbourhood mean of 3 by 3 model grid cells.

Download

The time series of average model–observation bias (MODEL-OBS) in H_s shown in Fig. 7 during each experiment period reflect a tendency for the global-scale wind driven UKW3g and UKC3owg simulations to under-predict significant wave heights across all months considered by up to 0.2 m on average. The sensitivity to coupling is found to be generally consistent across the different months. The increase in wave heights through enhanced winds in the fully coupled UKC3aow system tends to improve the bias for some periods during each month. However, there are as many periods when UKC3aow results become biased high relative to observations. On average, the impact of representing ocean–wave feedback processes in UKC3owg is shown to be relatively small (within 0.05 m) in comparison with the wind-related UKC3aow differences, with no clear improvement or degradation in performance across the experiments considered.

Figure 5 also shows the impact of coupling on peak wave period during the October 2014 experiment. Results are particularly improved along the south English coast for both UKC3aow (Fig. 5f) and UKC3owg (Fig. 5l). This is consistent with the case study results of Lewis et al. (2018a), and with Palmer and Saulter (2016). They found that the inclusion of surface currents in the Met Office UK4 operational wave system improved the representation of swell in this region. This was largely due to the refraction of long-period waves towards the coast due to wave–current interaction (evident in Fig. 5c and i). The reduction in quality of peak wave period results against observations along the south-eastern English coast in Fig. 5f is not apparent for UKC3owg (Fig. 5l), suggesting that the wind speed errors continue to dominate here.

Figure 8Sensitivity of wave model peak wave period (T_p) to coupling. (a, c, e, g) Time series of average model–observation (MODEL–OBS) bias across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to UKW3g. Note that plots have different scales across the different months evaluated, and observations are compared with a nearest neighbourhood mean of 3 by 3 model grid cells.

Download

Figure 8 shows the time series of model bias of peak wave period through each experiment. This highlights all model results producing waves that are of a longer period than observed at near-coastal sites for much of the time. Both the fully coupled UKC3aow and partially coupled UKC3owg simulations provide reduced biases.

4.5 Atmosphere component results

Comparing the UKC3aow and UKC3ao atmosphere results with the UKA3u atmosphere-only control simulation provides a strong test of the system. Whereas the coupled system SST evolves according to the free-running NEMO ocean model component, the surface forcing in the UKA3u control simulation has a daily-updating analysis-based (OSTIA; Donlon et al., 2012) SST in closer agreement with observations on average. The key limitation in UKA3u is therefore that the system has no information on the diurnal cycle of SST.

Figure 9Sensitivity of atmosphere model surface temperature to coupling. (a, d, g, j) Monthly mean difference between fully coupled and daily-updated OSTIA (UKC3aow–UKA3u) during April, July, October 2014 and February 2015 runs, respectively. (b, e, h, k) Monthly mean difference between fully coupled and partially coupled (UKC3aow–UKC3ao) runs during each experiment. (c, f, i, l) Percentage difference in surface temperature RMSE statistic for UKC3aow relative to UKA3u.

Download

Figure 9 summarizes differences in the surface temperature across both land and sea in UKC3aow fully coupled and UKC3ao partially coupled simulations relative to UKA3u. The distribution of monthly mean differences are quite varied between each month considered, reflecting both the seasonal variability in the quality of the coupled system ocean initial condition and its subsequent evolution (see Sect. 4.3). In April, October 2014 and February 2015 experiments, the coupled model SST tends to be warmer than OSTIA, while in July 2014 SST tend to be cooler away from the permanently mixed southern North Sea region and in the Celtic Sea. The impact of wave coupling processes on the SST evolution in UKC3aow relative to UKC3ao is as shown in Fig. 4 and discussed in Sect. 4.3. The comparison of results with in situ observations in Fig. 9 shows that in general the persisted SST from OSTIA better matches observations across much of the domain, as might be anticipated. However, notable areas where the UKC3aow surface temperatures are improved relative to OSTIA can be seen along at least some coastal regions in each month considered, where it is known that the satellite-based analysis product used in UKA3u is likely to be degraded by proximity to the coastline.

Figure 10(a, c, e, g) Time series of average model–observation (MODEL–OBS) bias of surface temperature across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to UKA3u. A negative relative |average bias| indicates the coupled system to have a lower average absolute bias across all observation sites than the control. Note that plots have different scales across the different months evaluated, and observations are compared with a nearest neighbourhood mean of 3 by 3 model grid cells.

Download

Noting the relatively large number of near-coastal sites contributing to the assessment, the time series of average bias during each month in Fig. 10 also show reasonably close agreement (typically within 0.5 K; slightly lower for the fully coupled UKC3aow case) of the coupled SST results with in situ observations, and extended periods of time when results show improved SST in the coupled simulations relative to UKA3u. The results for July 2014 in particular also highlight a diurnal cycle of SST bias in UKA3u, reflecting the daily persisted SST. In contrast, the representative dynamical representation of the diurnal SST cycle in the NEMO ocean model component results in a relatively smooth variation in the average bias for UKC3aow and UKC3ao.

Figure 11Sensitivity of atmosphere model air temperature at 1.5 m above surface to coupling. (a, d, g, j) Monthly mean difference between fully coupled and daily-updated OSTIA (UKC3aow–UKA3u) during April, July, October 2014 and February 2015 runs, respectively. (b, e, h, k) Monthly mean difference between fully coupled and partially coupled (UKC3aow–UKC3ao) runs during each experiment. (c, f, i, l) Percentage difference in air temperature RMSE statistic for UKC3aow relative to UKA3u.

Download

The key question to address is on the extent to which these differences in the surface temperature forcing, along with resulting differences in the surface momentum budget, change the coupled system meteorology relative to UKA3u. The spatial distribution of monthly mean differences in air temperature at 1.5 m above the surface due to coupling shown in Fig. 11 closely reflect the distribution of differences in mean surface temperature (Fig. 9). Differences over the ocean are dominated by the bias between the coupled ocean simulation and OSTIA, with the contribution due to diurnal cycle differences masked in the monthly mean measures presented. Differences over land are thought to result from a combination of advection of relatively warmer (cooler) air from over a nearby warmer (cooler) ocean between simulations, and of resulting differences to boundary layer and cloud development (e.g. Fallmann et al., 2017). Figure 11 also shows a general degradation of the agreement with in situ observations of 1.5 m air temperature over much of the domain, consistent with the SST results. However, specific regions can be seen where the fully coupled UKC3aow results have reduced RMSE in each month. The time series of average bias in air temperature across all observation sites shown in Fig. 12 is dominated by errors over land, which demonstrates a clear diurnal signal of the bias in all simulations (e.g. Bush et al., 2019). Simulated temperatures are relatively too cool during daytime and too warm at night. The impact of coupling during the periods considered in this study is, in general, to consistently shift the bias (typically warmed) at all times of the day, such that the bias is apparently “improved” relative to UKA3u during daytime but degraded at nighttime when UKA3u has a warm bias. Longer periods of improved air temperature results from UKC3aow are apparent during April, October and February experiments.

Figure 12(a, c, e, g) Time series of average model–observation (MODEL–OBS) bias of 1.5 m air temperature across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to UKA3u. A negative relative |average bias| indicates the coupled system to have a lower average absolute bias across all observation sites than the control. Note that plots have different scales across the different months evaluated, and observations are compared with a nearest neighbourhood mean of 3 by 3 model grid cells.

Download

Figure 13Sensitivity of atmosphere model wind speed at 10 m above surface to coupling. (a, d, g, j) Monthly mean difference between fully coupled and daily-updated OSTIA (UKC3aow–UKA3u) during April, July, October 2014 and February 2015 runs, respectively. (b, e, h, k) Monthly mean difference between fully coupled and partially coupled (UKC3aow–UKC3ao) runs during each experiment. (c, f, i, l) Percentage difference in wind speed RMSE statistic for UKC3aow relative to UKA3u.

Download

The distribution of experiment-mean 10 m wind speed changes due to model coupling are presented in Fig. 13. The differences between UKC3aow and UKC3ao highlight that the main impact of wave coupling across all months is a reduction in the monthly mean wind speed by up to 0.5 m s⁻¹. This is consistent with the wave-model-computed Charnock parameter tending to exceed the default assumed value of 0.011 across much of the domain (e.g. Fig. 2). This effect is most prominent during stormier periods, such as the October 2014 experiment.

Figure 13 shows a more varied spatial distribution of differences between UKC3aow and UKA3u, due to a more substantial difference in SST between the simulations, in addition to the Charnock parameter coupling. Comparing Fig. 13 with Fig. 9 suggests that to first order approximation, areas of relatively increased winds align with regions of relatively increased sea surface (and air) temperatures, and those with reduced wind speeds align with regions of reduced temperatures – i.e. the mean SST and wind speed anomalies are positively correlated. A growing literature has developed over recent years on the extent to which ocean temperature deviations drive atmospheric responses or vice versa at mesoscales in different regions of the world (e.g. Small et al., 2008; Gemmrich and Monahan, 2018). Figure 14 provides an initial assessment of the variability in monthly mean wind speed differences between simulations over the sea with differences in the near-surface temperature gradient (estimated as T_air−SST) and with differences in surface currents. This shows that where mean (land and sea) surface temperatures are increased (decreased) in UKC3aow relative to UKA3u, the impact tends to be a reduction (increase) in the near-surface temperature gradient. According to surface layer theory, under increasingly unstable conditions (change in stability < 0), the surface drag is increased and the near-surface wind speed is expected to increase (change in wind speed > 0). This mechanism is at least partly demonstrated by the variability of mean wind speed and near-surface temperature gradient differences shown in Fig. 14b. This contrasts with no clear relationship evident between wind speed and surface current differences in Fig. 14c. Corresponding results for UKC3aow and UKC3ao differences in Fig. 14 also show no strong dependencies between variables, highlighting the wind speed differences to be largely driven by the use of the wave model Charnock parameter in UKC3aow.

Figure 14Scatter plots showing relationships between differences in monthly mean results during July 2014 of (a, d) near-surface temperature difference (1.5 m air temperature − SST) with differences in SST, (b, e) 10 m wind speed with near-surface temperature difference and (c, f) 10 m wind speed with differences in surface current speed. Plots in (a)–(c) compare mean UKC3aow and UKA3u differences (i.e. fully coupled relative to atmosphere-only simulation). Plots in (d)–(f) compare mean UKC3aow and UKC3ao differences (i.e. fully coupled with wave relative to partially coupled atmosphere–ocean coupled).

Download

Figure 15(a, c, e, g) Time series of average model–observation (MODEL–OBS) bias of 10 m wind speed across all observing sites for each simulation for July, April, October 2014 and February 2015, respectively. (b, d, f, h) Differences in absolute |MODEL–OBS| bias for each simulation period relative to UKA3u. A negative relative |average bias| indicates the coupled system to have a lower average absolute bias across all observation sites than the control.

Download

Table 11Summary of computer resource usage and typical runtimes and volume of data outputs generated for each day of simulation of UKC3 systems. Note further optimizations of system node usage and runtimes are possible.

Download Print Version | Download XLSX

In common with air temperature results in Fig. 12, the time series of wind speed model bias for all simulations shown in Fig. 15 also demonstrate a diurnal cycle across all configurations, with winds too strong by up to 1 m s⁻¹ on average during daytime and slightly too weak at night. As an average across all sites, the relative impact of coupling is less pronounced than found for temperature variables, with changes due to ocean–atmosphere coupling within 0.1 m s⁻¹ for much of the periods considered. The impact of Charnock coupling is evident with improved results for UKC3aow relative to UKC3ao, and extended periods where the UKC3aow winds are improved relative to UKA3u, despite the improved SST specification in UKA3u. Results are consistently improved for UKC3aow during the February experiment for example. In contrast, from 15 October 2014, Fig. 15 shows both UKC3ao and UKC3aow simulations are degraded relative to the UKA3u control, coinciding with a period of relatively poorer surface temperature bias.

Despite the strong test set by comparing UKC3aow and UKC3ao performance to the atmosphere-only UKA3u control, these results demonstrate that representative simulations of the atmosphere can be performed in fully coupled mode at convective scales. It is clear that inclusion of the two-way feedbacks between surface waves and both ocean and atmosphere components provides some benefit over only including ocean–atmosphere feedbacks. The impact of coupling on the atmospheric boundary layer and associated features such as cloud development and near-surface visibility are the subject of ongoing research, focussing on more specific case studies periods and regions of interest (e.g. Fallmann et al., 2019).

4.6 Computational resource

Table 11 summarizes the typical computational resource usage and runtimes for a day of simulation on the Met Office Cray XC40 for each configuration used. No system optimization has been performed for UKC3, relative to the UKC2 configuration, and further opportunities for system optimization remain that will be pursued as part of the future UKC4 development, in particular relating to efficiency of the wave coupling. Previous ocean modelling research (e.g. Brown et al., 2011; Reza Hashemi et al., 2015) has commented on the large cost increase occurred due to wave coupling. Table 11 indicates that the UKO3 and UKW3 ocean-only and wave-only models used a relatively small amount of computational resource to run a day of simulation (4 and 6 node hours, respectively), and that using 12 node hours for a day of ocean–wave coupled simulation represents a big jump. However, this is not considered to be prohibitively expensive for research applications. In contrast, the computational impact of coupling in UKC3aow and UKC3ao is small, given that the regional atmosphere component is a relatively expensive part of the system.

Improving predictability

Improving skill beyond the UKC3 capability presented in this paper will require two key developments. Firstly, a number of the parameterization and parameter assumptions embedded across the MetUM atmosphere, NEMO ocean and WAVEWATCH III wave model codes, and their configuration in the regional system, will need to be re-examined and challenged. The UKC3 configuration still represents a “first look” implementation, in which a number of the underpinning assumptions and parameter choices established with uncoupled mode forcing or boundary data continue to exist. For example, consistent treatment of the atmospheric surface layer momentum budget across all codes has been an area of focus through the implementation of the wave-modified drag feedbacks within UKC3. The extent to which coupled mode running of these components is working against existing tuning is to be assessed. The influence of wind forcing on the skill of the wave model is a key example. The coupling exchange frequency is a further area deserving further study and optimization, given that the hourly coupling used in the UKC3 experiments presented in this paper is long relative to most studies discussed in Sect. 1.

Secondly, it is appropriate to begin developing the regional coupled system in an assimilative context, in order to improve the initial condition errors inherent within the current experimental design, as seen in the ocean SST results here for example. This is also a key step towards consideration of such systems for operational applications. Wada and Kunii (2017) have recently demonstrated the successful application of a regional mesoscale strongly coupled atmosphere–ocean data assimilation, implementing a local ensemble transform Kalman filter, for a tropical cyclone case study. For the UK regional coupled system development, it is more likely that a weakly coupled assimilation approach will be followed building on experience of developing this system for global coupled Numerical Weather Prediction (NWP) applications at the Met Office (e.g. Lea et al., 2015). It is not immediately clear how moving to an assimilative framework will modify any sensitivity to coupling, if at all, and research is clearly required in this area. One hypothesis is that the observational constraint will reduce the relative impact of representing feedback processes in the system. Conversely, improving the background state through improving the physics representation may in fact enhance the impact of data assimilation within the system by reducing the background errors. Even assimilation of observations in only one component may result in beneficial impacts across other model components. One particularly beneficial development that will support future research in this area is the implementation of analogous operational UK variable-resolution (UKV) atmosphere (e.g. Bush et al., 2019) and AMM15 ocean configurations (e.g. Graham et al., 2018) for operational forecasting at the Met Office using common physics settings, domain extent and grid definition to those used in UKC3. These provide the potential for operational analyses and boundary conditions for application in future regional coupled research activities.

Towards integrated environmental prediction capability

The focus of this paper has been on the physical coupling processes across atmosphere, ocean and wave components, with the driver of improving predictability through improved physical process representation. The vision for regional coupled prediction systems is that they can also provide a framework in which to represent interactions and feedbacks between physical, hydrological and biogeochemical cycles and processes at kilometre-scale. The initial focus in this context remains to demonstrate coupled prediction of the water cycle across atmosphere, land and ocean. Improvements to the convective rainfall representation and improved accuracy of quantitative precipitation forecasts in the UKC3 (RA1 physics) configuration relative to that in the UKC2 system (Bush et al., 2019) forms a key foundation of turning that potential into a reality. Further improvements to the hydrological capability in the JULES land surface model are also planned (e.g. Martínez-de la Torre et al., 2018), and their impact will be documented in the context of developing the UKC4 system.

Feedbacks between physical and biogeochemical processes in the ocean will also be introduced. A first required step is the technical implementation of the ERSEM (European Regional Seas Ecosystem Model; Butenschön et al., 2016) marine biogeochemical model as a new coupled component into the UKC4 system.

Longer-term considerations

The potential for delivering consistent natural hazard warnings across the scope of atmosphere, land, ocean and wave components was introduced in Sect. 1. Application of the UKC3 and its subsequent iterations to demonstrating this concept is still required. A key part of realizing this vision in a more operational context, particularly for hydrological and surge hazards for example, will be the requirement to develop an ensemble of regional coupled predictions. Consideration will need to be given on how to generate a regional ensemble with adequate spread in ocean and land surface states in the short term, with opportunities again to build on experience from the development of global-scale coupled NWP (e.g. Tennant and Beare, 2014). Incremental improvements to the capability and application of the UK regional coupled system will therefore continue over coming years, with the UKC3 system configuration providing an important milestone along that research journey. This effort will continue to require a multi-disciplinary approach, working in open collaboration both in the UK and with other groups around the world.

5.1 Code availability

Data availability

The nature of the 4-D data generated in running the various UKC3 experiments at 1.5 km resolution requires a large tape storage facility. These data are of the order of tens of terabytes. However, these data can be made available after contacting the authors. Each simulation namelist and input data are also archived under configuration management, and can be made available to researchers to promote collaboration upon contacting the authors.

Ocean bathymetry was obtained from the EMODnet Portal: EMODnet Bathymetry Consortium, EMODnet Digital Bathymetry (DTM), EMODnet Bathymetry (September 2015 release).