The Unified Model (UM; Davies et al., 2005) is the grid-point atmospheric model used in BARRA-R and ACCESS. It uses a non-hydrostatic, fully compressible, deep-atmosphere formulation and its dynamical core (Even Newer Dynamics for General atmospheric modelling of the environment, ENDGame) solves the equations of motion using mass-conserving, semi-implicit, semi-Lagrangian, time-integration methods (Wood et al., 2014). The model includes a comprehensive set of parameterizations, including a modified boundary layer scheme based on Lock et al. (2000), a variant of Wilson and Ballard (1999) for mixed-phase cloud microphysics, the mass flux convection scheme of Gregory and Rowntree (1990), and the radiation scheme of Edwards and Slingo (1996), which have all since been improved. Other parameterized sub-grid-scale processes include, fractional cloud cover and orographic drag. More details on all of the physics schemes can be found in Walters et al. (2017a).

The prognostic variables are three-dimensional wind components, virtual dry potential temperature and Exner pressure, dry density, and mixing ratios of moist quantities. The model is discretized on a horizontally staggered Arakawa C grid (Arakawa and Lamb, 1977) and a vertically staggered Charney–Phillips grid (Charney and Phillips, 1953). The staggered arrangement of grid points allows for accurate finite differencing but results in different model fields located on staggered grids displaced by half a grid of spacing along both axes. Data have been left on the staggered grids to allow users to apply the most appropriate re-gridding methods suited for given applications. The vertical levels smoothly transition from terrain-following coordinates near the surface to constant height surfaces in the upper atmosphere (Davies et al., 2005).

BARRA-R uses version 10.2 of the UM and is configured with 70 vertical levels extending from near the surface to 80 km above sea level: 50 model levels below 18 km and 20 levels above this. While configured with this height based on ACCESS-R, we have more confidence in the data up to a height of 25–30 km where we have the most information from observations. The horizontal domain of BARRA-R spans from 65.0 to 196.9∘ E and -65.0 to 19.4∘ N (Fig. 1), with constant latitude and longitude increments of 0.11∘ × 0.11∘ and 1200 × 768 grid points in the horizontal. Our choice of the horizontal resolution follows the deterministic component of the UKMO reanalysis and the IMDAA reanalyses. The model was run to produce 12 h (12 h) forecasts in each 6-hourly cycle (see Sect. 2.2) to give extra data for driving dynamical downscaling within the domain.

The model parameterizations in BARRA-R are inherited from the UKMO Global Atmosphere (GA) 6.0 configurations described in Walters et al. (2017a). The GA6 configurations are also suited for limited-area models with resolutions >10 km but with some modifications, as described below. i.

A variable Charnock coefficient is used in surface heat exchange over the sea to improve the tropical Pacific air–sea exchange (Ma et al., 2015).

The BARRA-R analysis scheme is based on fixed deterministic atmospheric and land-surface assimilation systems used by the UKMO in UERRA (Jermey and Renshaw, 2016) and IMDAA (Mahmood et al., 2018). BARRA-R uses a sequential data assimilation scheme, advancing forward in time using 6-hourly analysis cycles centred at synoptic hours t0 = 00:00, 06:00, 12:00, and 18:00 UTC and 12 h forecast cycles from t0-3 h (Fig. 2).

In each analysis cycle, available observations, distributed across a 6h analysis window t0-3 h ≤ t<t0+3 h, are combined with the prior information of the model forecast from the previous cycle (the background state) to provide a more accurate estimate of the atmosphere over this window. This first involves a four-dimensional variational (4D-Var) analysis of the basic upper-air atmospheric fields (wind, temperature, specific humidity, pressure) with conventional and satellite observations (see below). The 4D-Var is favoured over 3D-Var as it takes account of time tendency information in the observations and this has a positive impact on the resulting forecasts (Rawlins et al., 2007). The UKMO's VAR assimilation system (version 2016.03.0) is used. The 4D-Var uses a linear perturbation forecast (PF) model (Lorenc, 2003; Rawlins et al., 2007; Lorenc and Payne, 2007), which uses a simpler model state linearized about a “guess” trajectory (i.e. tangent linear model) with a lower resolution (0.33∘ cf. 0.11∘) than the full forecast model. The lower resolution is chosen to limit computational costs. The PF model uses a simplified set of physical parameterizations including a simple boundary layer, cloud latent heat release, large-scale precipitation, and convection. In other words, it is assumed that the lower-resolution corrections to the background state (i.e. increments), interpolated to a higher resolution, are suitable corrections for the full model. The analysis increments from 4D-Var valid at t0-3 h are added to the background state at t0-3 h to produce an improved initial condition for the forecast model to perform the next 12 h forecast from t0-3 h to t0+9 h. A constraint of zero analysis increments is specified at the model boundary such that BARRA-R relies on the driving model ERA-Interim to define large-scale flow and other atmospheric conditions (Sect. 2.1.3). The observation departure statistics of the analysis, which are differences between the analysis and observations, are shown to be less than those of the model background (Supplement, Table S1). The assimilation is therefore behaving as desired by drawing the model towards observations for nearly all observational types.

The variational method minimizes a cost function whose two principal terms penalize distance to the background state and distance to the observations. The two terms are squared differences weighted by the inverse of their corresponding error covariances. In BARRA-R, the background error covariance has been estimated by a smooth parameterized approximation to climatology tuned by forecast differences (Ingleby, 2001). Accordingly, the estimated background error covariance is invariant between successive analysis windows but is time varying within the analysis window. The cost function also includes a pressure-based energy norm that serves as a weak-constraint digital filter to suppress spurious fast oscillations associated with gravity-inertia waves produced in model forecasts when analysis increments are added to the background state (Gauthier and Thépaut, 2001).

The initial land-surface state can have a significant impact on short-term forecasts of screen-level temperature and humidity, and its quality can also be improved through data assimilation. An extended Kalman filter (EKF) using observations of 2 m temperature and humidity is used to analyse the BARRA land state at every 6 h cycle and provide analyses of soil moisture, soil temperature, and skin temperature as described by Dharssi et al. (2012). The assimilation of satellite-retrieved soil moisture is not attempted here as it has not been realized in ACCESS. The UKMO's SURF analysis system (version 2016.07.0) is used to perform EKF. The Jacobian, which relates observed variables to model variables, for the Kalman gain matrix is estimated using finite difference by perturbing each model variable to be analysed in 40 perturbations and performing short 3 h forecasts. Here JULES (version 3.0) is run in the stand-alone mode, decoupled from the UM. The BARRA-R land state is reconfigured with EKF-derived surface analyses at every t0.

Note that the last 6 h forecast of a model run represents the prior state estimates needed for the next analysis cycle. The forecast fields valid at t0-3 h, t0-2 h, and t0-1 h are discarded, as these fields may still be influenced by transient artefacts due to the slight imbalance introduced by the addition of the analysis increments. It is already noted that this effect is also mitigated with the energy norm in the 4D-Var's cost function that penalizes the unbalanced structure in the increments.

The reanalysis is produced with multiple parallel production streams to speed up production. Each stream has a month of spin-up time from the ERA-Interim initial conditions before production data are archived, with most streams producing 1 year of reanalyses. Trials have shown that a 1-month period is a sufficient spin-up for the atmosphere (Renshaw et al., 2013) and top levels of soil moisture but insufficient for soil moisture in the deeper layers.

Conventional observations from land-surface stations, ships, drifting buoys, aircrafts, radiosondes, wind profilers, and satellite observations, namely retrieved wind, radiances, and bending angle, are assimilated in BARRA-R. The various observational types are chosen as they have been assimilated in the Bureau's operational NWP systems; other observational types, such as clear-sky radiances, have not been assimilated due to resource constraints. Rain observations from radar and gauges are also not assimilated as their assimilation schemes are still being tested for operational NWP. As listed in Table 1, the datasets are pragmatically taken from multiple sources, as they are being prepared during the production runs. Most of the observations prior to 2009 are supplied by the ECMWF, and the satellite radiance data from 2017 and onwards are extracted from the UKMO operational archive.

The Bureau's archived observational data are also used to support this work, especially for the cycles from 2010 onwards. BARRA-R also assimilates additional high-frequency (10 min) land-surface observations from automatic weather stations in Australia and locally derived satellite atmospheric motion vectors (AMV). Ground positioning system (GPS) radio occultation bending angle data up to 2009 are provided by the Radio Occultation Meteorology Satellite Application Facility (ROM SAF). Additional land-surface observations over New Zealand are extracted from their National Climate Database (CliFlo, 2017). The 4D-Var assimilation of local AMV (Le Marshall et al., 2013) and GPSRO (Le Marshall et al., 2010) has been shown to improve operational forecasts.

Before being assimilated, observations are screened to select the best-quality observations, remove duplicates, and reduce data redundancy via thinning using the UKMO's Observing Processing System (OPS; based on version 2016.03.0) (Rawlins et al., 2007). There are per-cycle quality controls performed based on the method of Lorenc and Hammon (1988). Observations significantly different from the model background are rejected when exceeding a threshold calculated by a Bayesian scheme unless they are consistent with other observations nearby. The observational error variances and thinning distances are established at the UKMO and the Bureau for their NWP systems. For the surface, sonde, and aircraft observations, an observation automatic monitoring system performs monthly blacklisting of sites that show consistently large differences with BARRA-R's forecast over a 1-month period. The system also calculates bias corrections for surface pressure and for aircraft and sonde temperature.

For the satellite data, instruments and their individual channels are rejected when they become unreliable. The blacklisting is informed by the work of the ECMWF and MERRA-2 reanalysis teams. Further, air-mass-dependent variational bias correction is applied to satellite radiances as part of the assimilation process, allowing the time-varying corrections to fit drifts in instrumental bias (Harris and Kelly, 2001; Dee and Uppala, 2009). The bias corrections were calculated monthly, with the satellite radiances during the first month of each production stream not assimilated. There are abrupt changes to the amount of satellite data assimilated at the start and end of satellite missions and the various observational data archives. In some cases, changes occur when corrections were made to the observation screening and thinning rules mid-production for the 2010–2015 reanalyses. The impacts of such changes, known to cause artificial shifts and spurious trends in a reanalysis (e.g. Thorne and Vose, 2010; Dee et al., 2011), are still to be investigated for BARRA-R.

To assess BARRA-R in the atmosphere, we compare the t0+6 h forecasts on pressure levels with radiosonde and pilot wind observations at 00:00 and 12:00 UTC on standard pressure levels ranging from 1000 to 10 hPa, using the harmonized dataset produced by Ramella Pralungo et al. (2014a, b). The pressure-level fields of BARRA-R and ERA-Interim analyses at time t0 are also compared, even though they are not independent from the observations; such comparisons only provide baselines to interpret the relative quality of the BARRA-R forecasts. Similar comparisons with ERA-Interim's twice-daily forecasts at these observation times are also not possible because they start from 00:00 and 12:00 UTC. The model data are interpolated horizontally to the sonde and pilot launch locations via bilinear interpolation, and RMSD is calculated at each location and pressure level. The resulting box plots of RMSD are shown in Fig. 7. Depending on the pressure level and parameter evaluated, between 54 and 203 sites were available. There is a marked variability in RMSD with the pressure levels, particularly for wind speed, due to factors such as variations in the number of observing sites, increasing sonde drift error on ascent, and differences in the dynamic range of the fields with height. A markedly higher RMSD for wind speed occurs at 200 hPa, a height a which the jet stream can be located.

It is difficult to discern the differences between the two analyses, suggesting that they perform similarly from assimilating the same observations. Assimilation at a coarser resolution of 0.33∘ (cf. 0.11∘ of the forecast model) in BARRA-R does not drastically improve 0.75∘ representations of temperature and wind at these pressure levels and at point scales. There are also small differences between the analyses and BARRA-R background, indicating that the 0.11∘ forecast model does not degrade from the lower-resolution analysis of BARRA-R but also does not improve upon the ERA-Interim 0.75∘ representation of these fields at the observation locations.

Figure 8 compares BARRA-R's 00:00 UTC analysis of air temperature at 850, 700, and 500 hPa against the analyses from ERA-Interim and MERRA-2 (M2I3NPASM). BARRA-R is cooler at 500 hPa across the domain and warmer at 850 hPa in the tropics than the global reanalyses, and the monthly differences in the zonal mean are of the order of 1 K. BARRA-R also shows a cooling shift at 700 and 500 hPa in the tropics and a warming shift south of 40∘ S after 2010. But when compared against MERRA-2, in the tropics, BARRA-R is warmer at 700 hPa, and the apparent shift in BARRA-R is also seen in MERRA-2 (relative to ERA-Interim) at these levels.

We consider three reference gridded datasets to compare with the reanalyses. First is the 0.05∘ × 0.05∘ rain gauge analysis of daily accumulation over Australia from AWAP, produced using the Barnes method in which the ratio of observed rainfall to monthly average is used in the analysis process (Jones et al., 2009). There is a north–south gradient in the AWAP analysis errors with larger analysis errors in the northern tropical regions, where the length scales of convective rainfall events are shorter and more variable (Jones et al., 2009). Second is the 1∘ × 1∘ (full data daily) rain gauge analysis over the domain from the Global Precipitation Climatology Centre (GPCC version 2018; Ziese et al., 2018), created using an empirical weighting-based interpolation method described in Becker et al. (2013). As with AWAP, GPCC is less accurate in regions where station scarcity and high precipitation variability coexist. For instance, different GPCC interpolation methods can yield very different analyses over the south Asia region (Becker et al., 2013). The third reference is the 0.25∘ × 0.25∘ satellite-based analysis of 3-hourly rain rates from the Tropical Rainfall Measuring Mission (TRMM) multi-satellite precipitation analysis (TMPA 3B42 version 7; Huffman et al., 2006). TMPA 3B42 combines precipitation estimates from various satellite systems and rain gauge monthly analysis. Satellite-derived estimates of convective precipitation are largely accurate in the low latitudes (Ebert et al., 2007; Chen et al., 2013), but the TMPA product is less accurate over the ocean due to the absence of local observations used for gauge adjustments (Sapiano and Arkin, 2009) and south of 40∘ S due to limited local cross-sensor calibration (Huffman et al., 2006). TRMM often underestimates precipitation in high-latitude regions with significant topography due to difficulties of satellite retrievals over snow-covered surfaces and/or due to the high elevations (Barros et al., 2006; Matthews et al., 2013). TRMM is also known to underestimate light rainfall and drizzle over subtropical and high-latitude oceans (Berg et al., 2010). In addition to these considerations, there are inherent limitations in comparing the reanalyses with AWAP, GPCC, and TMPA. Specifically, products with coarser grids tend to over-represent low-threshold events occurring at spatial scales smaller than their grid sizes and under-represent high-threshold events. Further evaluation of BARRA-R precipitation estimates against point gauge observations and AWAP are reported in Acharya et al. (2019).

Neither BARRA-R nor ERA-Interim assimilated rainfall observations. Precipitation estimation from their forecast models is constrained by other observation types. Following Sect. 2.1, in BARRA-R, the microphysics scheme based on Wilson and Ballard (1999) parameterizes the atmospheric processes that transfer water between the four modelled states of water (vapour, liquid droplets, ice, and raindrops) to remove moisture resolved on the grid scale. As the 12 km model is not “storm resolving”, BARRA-R uses the mass flux convective parameterization scheme of Gregory and Rowntree (1990) with the CAPE closure to model sub-grid-scale precipitating and non-precipitating convection using an ensemble of cumulus clouds as a single entraining–detraining plume. Such a scheme prevents the unstable growth of cloudy structures on the grid, which is otherwise required for explicit vertical circulations to develop (Clark et al., 2016). The modelled convection also works independently at each grid point, and the model can only predict the area-average rainfall instead of the spectrum of rainfall rates. Consequently, BARRA-R's precipitation estimates from sub-grid convection will be more erroneous than those for large-scale precipitation. In other words, the accuracy of BARRA-R is expected to worsen during the warm season and at low latitudes and to improve during the cooler season and at high latitudes where non-convective precipitation is dominant. To allow the UM to spin up from the analysis increments, we examine the quality of the precipitation accumulation between t0+3 h and t0+9 h by comparing against gridded datasets. This also addresses the issue that the UM yields excess precipitation at analysis time (t0-3 h) due to a temporary imbalance in the moisture fields by allowing time for the model to adjust and remove the excess. For ERA-Interim, we used its first 12 h accumulation, which is considered the most accurate (Kallberg, 2011).