The BRAMS original prognostic equation for the Exner function was derived by Klemp and Wilhelmson (1978, hereafter KW78). The prognostic equation was obtained by combining the ideal gas equation with the mass continuity equation for compressible fluids. Medvigy et al. (2005, hereafter M05) expanded the original Eq. (6), which now reads as ∂π′∂t=-Rπ0cvρ0θ0∂ρ0θ0u∂x+∂ρ0θ0v∂y+∂ρ0θ0w∂z︸heat  flux-u∂π′∂x-v∂π′∂y-w∂π′∂z︸advection-Rπ′cv∂u∂x+∂v∂y+∂w∂z︸divregence+R(π′+π0)cvθvdθvdt︸heating. KW78 pointed out that the first term of the right-hand side of Eq. (7) typically has a higher order of magnitude than the other terms in studies of cloud dynamics, and the simplified version of Eq. (7) became the standard solution in both RAMS and BRAMS. However, KW78 also pointed out that the simplified equation violates mass conservation and deteriorates the accuracy of predicted pressure fields. M05 evaluated the conservation of mass in a regional simulation for New England, found the loss rates to be as large as 3 % day-1, and showed a significant improvement when the full equation was included. In this version of BRAMS, both the native and complete forms of the prognostic equation are available, and following M05 implementation, in BRAMS 5.2 we also solved the advection, divergence, and heating terms of Eq. (7) using the main time step, whereas the heat flux term is updated using the acoustic time step.

RAMS employs a hybrid time integration scheme combining a leapfrog scheme for the wind components and an Exner function with forward-in-time for scalars. The computational mode produced by the leapfrog scheme is damped with the application of the Robert–Asselin time filter (Asselin, 1972), which makes the overall accuracy first order. Williams (2009) proposed a simple modification to this time filter with a few extra lines of coding but increasing the accuracy of the scheme to third order. This improved time filter is available in BRAMS by appropriate setting of a flag in the namelist input file (RAMSIN).

A third option for time integration in BRAMS was based on the work of Wicker and Skamarock (2002, hereafter WS2002). This scheme has proven to be very robust and efficient, being applied in several state-of-the-art non-hydrostatic atmospheric models (e.g. Skamarock and Klemp, 2008; Baldauf, 2008, 2010; Skamarock et al., 2012). The WS2002 scheme is of a low-storage Runge–Kutta type with three stages and a third order for linear problems (hereafter RK3). The three stages require three evaluations of the slow mode tendencies (e.g. the advection term); however, this cost is offset by the larger time step allowed by the scheme when allied with a high-order advection scheme (see the discussion in Sect. 2.1.3).

The last option for a time integration scheme was described by Wicker (2009). This technique consists of a combination of two different schemes applied in two steps. A predictor step is performed by applying Adams–Bashforth of a second-order scheme and then a corrector step is completed by applying Adams–Moulton of a third-order scheme (hereafter ABM3). ABM3 is of third order and requires only two evaluations of the slow mode tendencies, demanding, however, a larger memory footprint than RK3 and a shorter time step. The advantage of using ABM3 over RK3 might arise when the length of the time step required by model stability is not dictated by the advective transport but by other physical processes (e.g. cloud microphysics).

An additional advection scheme, which preserves the initial monotonic characteristics of a scalar field being transported with simultaneously levying low numerical diffusion, is available in BRAMS. The method developed by Walcek (2000) is highly accurate and absolutely monotonic. Freitas et al. (2012) reported its implementation in BRAMS and related impacts on the accuracy of the transport of relatively inert tracers as well as on the formation of secondary species from nonlinear chemical reactions of precursors. The results revealed that the new scheme produces much more realistic transport patterns, without generating spurious oscillations and undershoots and overshoots or diffusing mass away from the local peaks. Besides these features, the scheme also presents good performance in retaining nonlinear tracer correlations and conserving the mass of multi-component chemical species. The latter feature is not evident since monotonic preserving filters typically make the numerical advection scheme non-strictly linear.

As an example of the application of this scheme within BRAMS, the advection of a hypothetical rectangular parallelepiped tracer field by a realistic 3-D wind flow is discussed as follows. The model was configured with one grid with 10 km horizontal grid spacing covering the southeastern part of Brazil and with a time step of 15 s. The total length of the time integration was 24 h. The tracer mass mixing ratio is initiated with 100 au and the background is set to zero. The horizontal domain initially occupied by the tracer is shown in panel a of Fig. 1, while in the vertical the tracer was initially localized between 1.7 and 4.1 km in height (not shown). The tracer mass mixing ratio distribution 12 h after and simulated by the original and monotonic advection schemes is shown in panels b and c of Fig. 1, respectively. In this study, the original advection scheme of BRAMS noticeably introduced spurious oscillations, overshoots, and undershoots (panel b), the latter with negative values of the mass mixing ratio (see Freitas et al., 2012, for further details). On the other hand, the simulation produced by the new scheme is much better at keeping the monotonicity of the distribution without spurious oscillations and negative mass mixing ratios (panel c), even for a real strongly divergent and deformational wind, as in this case.

Following WS2002, BRAMS also has a new set of advection schemes to be applied in conjunction with RK3 or ABM3 time schemes. The set is comprised of first- to sixth-order spatial approximations for the fluxes at the edge of the grid cells. Also, exactly the same flux approximation can be applied for advection of scalars and momentum. The positivity constraint for scalars can be applied following Skamarock (2006).

Future versions of BRAMS will also include monotonicity constraints for scalars and an option for the WENO (weighted essentially non-oscillatory) third- and fifth-order formulations (Baba and Takahashi, 2013) for the advection operators.

The current version of the two-moment (2M) microphysical parameterization used in RAMS, version 6, has been implemented in BRAMS. This scheme has prognostic equations for number concentration and mixing ratio for eight hydrometeor categories (cloud, drizzle, rain, pristine, snow, aggregates, graupel, and hail). Each hydrometeor size spectrum is described by a generalized gamma distribution with a user-specified shape parameter (Meyers et al., 1997; Saleeby and Cotton, 2004, 2008).

According to Cotton et al. (2003), the 2M microphysical scheme comes with an efficient and stable algorithm for heat and vapor diffusion without requiring numerical iteration (Walko et al., 2000), sea salt and dust treatment, and a bin sedimentation scheme. Lately, Saleeby and Cotton (2008) developed a binned approach to cloud-droplet rimming, which computes the collision–coalescence process between ice and cloud particles in a more realistic way.

Cloud and drizzle number concentrations are computed from cloud condensation nuclei (CCN) and giant CCN (GCCN) concentrations, respectively. A lookup table (LUT) is used to obtain the CCN concentration that is activated as a function of aerosol size, concentration, and composition via hygroscopicity parameter (Petters and Kreidenweis, 2007), as well as updraft velocities, pressure, and temperature. On the other hand, GCCN activation does not depend on the environmental conditions, being completely used in the drizzle nucleation process. Both aerosol categories may be advected, diffused, depleted, and restored (by droplet evaporation) as well as have their initial concentrations specified by the user as either homogeneous or heterogeneous fields (Saleeby and Cotton, 2004, 2008).

The aerosol aware bulk microphysics scheme described in Thompson et al. (2008) and Thompson and Eidhammer (2014), hereafter GT, was also implemented in BRAMS. The GT scheme treats five separate water species, mixing single- and double-moment treatment for different cloud species to minimize computational cost. It also includes the activation of aerosols as cloud condensation (CCN) and ice nuclei (IN) and, therefore, explicitly predicts the droplet number concentration of cloud water as well as the number concentrations of the two new aerosol variables, one each for CCN and IN. The aerosol species are lumped into two different groups according to their hygroscopicity. Hygroscopic aerosols are in the general category of “water friendly” (Nwfa), and the non-hygroscopic ice-nucleating aerosols are in the group “ice friendly” (Nifa). As a first approximation, Nifa is assumed to be only mineral dust in the accumulation mode, and all the other species (sulfates, sea salts, organic matter, and black carbon) are assumed to be a mixture of the species in each population and allocated to the hygroscopic mode Nwfa.

Aerosol activation also uses a LUT of activated fraction determined by temperature, vertical velocity, aerosol number concentration, and hygroscopicity parameter determined by the model. The LUT was built following Köhler activation theory within a parcel model from Feingold and Heymsfield (1992) with additional changes by Eidhammer et al. (2009) to use the hygroscopicity parameter (Petters and Kreidenweis, 2007). This approach is similar to the one used by RAMS CSU microphysics previously described (Saleeby and Cotton, 2004, 2008). However, the LUT of GT has a coarser variation in terms of hygroscopicity parameters compared to RAMS CSU. The coarse resolution of the LUT in terms of aerosol hygroscopicity contributes to the GT scheme's low cost, but also represents a limitation for the ambient with high loads of very low hygroscopic aerosols, such as biomass burning affected areas (Sánchez Gácita et al., 2016).

As a low-cost option for the explicit aerosol aware microphysics schemes described above, the parameterization of aerosol particle activation as CCN was also implemented following the approach of Abdul-Razzak and Ghan (2000, 2002). This scheme, in its form for multiple log-normal distributions, assumes that the particles are in equilibrium with the environment, and the terms of curvature and the solute in the particle growth after activation can be neglected. As a first approach, for applications in a black-carbon rich atmosphere, the aerosol activation can be done via Abdul-Rassack parameterization and feed either the GT or RAMS CSU microphysics directly with the CCN number concentration.

The BRAMS radiation module includes two additional schemes to treat atmospheric radiative transfer consistently for both longwave and shortwave spectra. The first scheme is a modified version of the Community Aerosol and Radiation Model for Atmospheres (CARMA) (Toon et al., 1989), and the second one is the Rapid Radiation Transfer Model (RRTM) version for GCMs (RRTMG, Mlawer et al., 1997; Iacono et al., 2008). RRTMG shares the same basic physics as RRTM, though it incorporates several modifications (Iacono et al., 2008) in order to improve computational efficiency. The CARMA and RRTMG schemes both solve the radiative transfer using the two-stream method and include all the major molecular absorbers (water vapor, carbon monoxide, ozone, oxygen) and aerosol extinction. The RRTMG implementation preserved all the absorption coefficients for molecular species used in the correlated k distribution method, which were based on a line-by-line model (Iacono et al., 2008). CARMA treats gaseous absorption coefficients using an exponential sum formulation (Toon et al., 1989).

Radiation schemes in BRAMS are both in-line coupled with the aerosol and cloud microphysics modules to provide online simulations of aerosol–cloud–radiation interactions. The CARMA and RRTMG radiative schemes are both fed with aerosol optical depth (AOD) profiles calculated from simulated aerosol mass loading and prescribed aerosol intensive optical properties, specifically the extinction efficiency, single scattering albedo, and asymmetry parameter taken from a LUT. Aerosol intensive optical parameters' prescription is regionally dependent. For South America, the parameters present in the LUT (Procopio et al., 2001; Rosário et al., 2013) are obtained from offline Mie calculations using as input climatological particle size distribution and the complex refractive index from sites of the AErosol RObotic NETwork (AERONET, Holben et al., 1998) distributed across South America. As an example, Figs. 2 and 3 present comparisons between BRAMS 5.2 simulation with the CARMA scheme for a set of diurnal cycles of downward shortwave and longwave irradiance, respectively, at the surface, with measurements at a pasture site (Abracos Hill – 10.760∘ S, 62.358∘ W) in the southern portion of the Amazon Basin. For both radiation spectrums, the model reproduced consistently the diurnal cycle of the surface downward radiative energy. In the case of the shortwave radiation, the inclusion of the aerosol radiative effect proves to be fundamental to modeling its diurnal cycle.

Cloud physical (ice and liquid water path and particle sizes) and optical properties (optical depth) in the CARMA radiative scheme have been parameterized according to Sun and Shine (1994), Savijärvi (1997), and Savijärvi et al. (1997, 1998) using liquid and ice water content profiles provided by the BRAMS cloud microphysical module. In this case, subgrid-scale cloud variability is not taken into account.

For the RRTMG scheme, the optical properties of liquid and ice water are from Hu and Stamnes (1993) and Ebert and Curry (1992), respectively, and sub-grid-scale cloud variability including cloud overlap is statistically addressed with MCICA (Iacono et al., 2008), the Monte Carlo independent column approximation (Barker et al., 2008; Pincus et al., 2003). The MCICA approach presupposes that cloud liquid water and ice, and cloud fraction, are prognostic variables. As such, the cloud liquid water effective radius was parameterized in BRAMS following the generalized power-law expression of Liu et al. (2008): rel=34πρw1/3βLWCN1/3 where LWC is the liquid water content, and is the water density, and N is the cloud droplet number concentration. L and N are in CGS units. β is a dimensionless parameter that depends on the spectral shape of the cloud droplet distribution, set based on observation as β=aβLWCN-bβ, with aβ and bβ equal to 0.07 and 0.14, respectively.

The cloud radiative forcing is very sensitive to the determination of β. According to Liu et al. (2008), β increases with aerosol loading and leads to a warming effect that acts to substantially offset the cooling of the Twomey effect by a factor of 10 to 80 %. A bβ>0 leads to a weaker dependence of rel on LWC /N and a smaller indirect aerosol effect, with a better agreement with observation. In principle, this generalized power-law expression for re effectively accounts for the increase in droplet concentration and the decrease in droplet size due to aerosol (Twomey, 1974), as well as the reduction in precipitation efficiency, which increases the liquid water content, the cloud lifetime (Albrecht, 1989), and the cloud thickness (Pincus and Baker, 1994).

The ice effective radius was parameterized in BRAMS following Wyser and Yang (1998), with an explicit dependence on both ice water content and temperature: rei=377.4+203.3B+37.91B2+2.3696B3,B=-2+10-3273-T1.5log⁡10IWCIWC0, where T is the temperature in Kelvin, IWC is the ice water content in gm-3, and IWC0=50 gm-3.

This parameterization assumes the ice crystals consisting of hexagonal columns, and so is compatible with the ice optical properties from Ebert and Curry (1992) assumed in RRTMG.

In addition, to fulfill MCICA requirements, a cloud fraction representation was also implemented in BRAMS, based on the parameterization originally from the Community Atmosphere Model (CAM, http://www.cesm.ucar.edu/models/cesm1.2/cam/), which is a generalization of the scheme introduced by Slingo (1987), with variations described in Kiehl et al. (1998), Hack et al. (1993), and Rasch and Kristjansson (1998). In this representation, three types of cloud are diagnosed, depending on relative humidity, atmospheric stability, and convective mass fluxes: low-level marine stratus, shallow and deep convective clouds, and layered cloud.

The marine stratus clouds are located according to the identification of stable layers between the surface and 700 mb (> 0.125 K mb-1) and using the following empirical relationship from Klein and Hartmann (1993): Cst=min⁡1.0,max⁡0.0,θ700-θs0.0057-0.5573, where θ700 and θs are the potential temperatures at the 700 mb and surface levels, respectively. The stratus clouds are located just below the strongest stability jump between these two levels.

The convective cloud fraction follows a formulation based on the updraft mass flux, both for shallow and deep (Xu and Krueger, 1991; Xu and Randall, 1996): Cshallow=k1,shallowln⁡(1.0+k2Mc,shallow),Cdeep=k1,deepln⁡(1.0+k2Mc,deep), with k1,shallow=0.07, k1,deep=0.14, and k2=500, and Mc is the convective mass flux at the given level.

Any other clouds are diagnosed according to the relative humidity: Cc=min⁡0.999,max⁡Rh-RHmin1-RHmin2,RHmin=RHminlow,p≥750mb,RHminhigh,p<750mb, with RHminlow=0.90 and 0.80, over water and land, respectively, and RHminhigh=0.80. Cc=min⁡0.999,max⁡Rh-RHmin1-RHmin2 The total cloud fraction in the grid cell is Ctot=min⁡1,max⁡Cst,Cdeep,Cshallow,Cs. The total cloud optical depth is given by the contribution from liquid and ice water contents, and accounts for the cloud fraction.

In BRAMS, as in the original RAMS formulation, the local changes in momentum and scalars due to turbulent transport depend on the divergence of turbulent fluxes (RAMS, 2003). When the grid resolution is coarser than the size of the largest eddies (typically coarser than 100 m–1 km), the eddy covariance fields needed to determine the turbulent fluxes are determined through K-theory (Stull, 1988), which requires the determination of eddy diffusivities for momentum and scalar quantities u′v′‾=v′u′‾=-Kmh∂u∂y+∂v∂x,uh′w′‾=-Kmz∂uh∂z,w′uh′‾=-Kmz∂w∂xh,uh′ε′‾=-Khh∂ε∂xh,w′ε′‾=-Khz∂ε∂z, where (x,y) are the horizontal directions (xh is either x or y), z is the vertical direction, (u, v) are the horizontal wind directions (uh is either u or v), w is the vertical velocity and ε is any scalar, Kmh and Kmz are the horizontal and vertical diffusivity coefficients for momentum, and Khh and Khz are the horizontal and vertical diffusivity coefficients for scalars. It is important to note that Eqs. (3)–(4) are only to be used when the grid horizontal resolution is much coarser than the vertical resolution because they violate vorticity conservation; however, different scales are needed when the horizontal and vertical grid resolutions are different to avoid numerical instabilities (RAMS, 2003).

The horizontal diffusivities are determined using the same algorithm implemented in RAMS, which is based on Smagorinsky (1963), with the inclusion of the Brunt–Väisäla correction by Hill (1974). For the vertical diffusivity coefficients, we use a vertical parameterization based on the level-2.5 model by Mellor and Yamada (1982), further modified by Nakanishi and Niino (2004). In this model, the diffusivity coefficients depend on turbulent kinetic energy per unit mass (TKE), which also becomes a prognostic variable: Kmz=LqSm,Khz=LqSh,q=2TKE=u′u′‾+v′v′‾+w′w′‾, where L is the master length scale and Sm and Sh are non-dimensional stability functions for momentum and buoyancy. Both L and the stability functions are determined following Nakanishi and Niino (2004), which allows for stronger turbulence and deeper boundary layers compared to the original formulation. The non-dimensional stability functions also include a correction factor to avoid numerical instabilities under growing turbulence (see Helfand and Labraga, 1988) and an upper limit on L under very stable conditions to avoid TKE becoming negative, following the implementation by Janjić (2001) and Nakanishi and Niino (2006). Although Nakanishi and Niino (2004) also described a higher-order, level-3 parameterization, this would require including prognostic equations for the variance of every scalar and the covariance between pairs of scalars, which would rapidly become unmanageable due to high computational load (Mellor and Yamada, 1982).

BRAMS also offers the possibility of using a combination of the LEAF surface scheme (Walko et al., 2000) and the Town Energy Budget (TEB) (Masson, 2000; Freitas et al., 2007). Use of the bare soil formulation or the adjustment in the surface–vegetation–atmosphere transfer (SVAT) scheme parameters is very frequent. However, as stressed by Masson (2000), such an approximation is satisfactory only for large temporal or spatial averages, and it is necessary to incorporate a more detailed scheme when smaller scales are considered. Therefore, the simulation of several mesoscale and local processes that simultaneously occur in an urban atmosphere and its surroundings requires a more detailed urban surface parameterization. Such processes include the circulations generated by an urban heat island (UHI) and its interaction with other atmospheric phenomena (Freitas et al., 2007; Nair et al., 2004), air pollution (Andrade et al., 2004; E. D. Freitas et al., 2005), and human comfort conditions (Johansson et al., 2013), among others. In BRAMS 5.2, the TEB and LEAF schemes are activated simultaneously, and the surface fluxes of momentum and moisture, temperature, surface albedo, and emissivity are calculated by TEB wherever an urban grid point is identified, while LEAF is applied elsewhere (e.g. bare soil, water bodies, grass, forest, or any vegetation). TEB considers the interaction of shortwave and longwave radiation with the urban structure, allowing multiple reflections with walls and roads. In addition to the 3-D urban structure in the TEB formulation, another advantage is the possibility of simulating anthropogenic heat and moisture fluxes emitted both by mobile sources, such as heavy and light duty vehicles, and fixed sources, such as industries, commerce, and domestic activities in general. For large cities, such as São Paulo and Rio de Janeiro, the anthropogenic heat sources are key features, not only for meteorological reasons, but also for health and public policy management. As anthropogenic contributions can vary strongly depending on the urban area, the implementation of TEB in BRAMS allows the user to define those contributions in the model configuration file. Following the work of Khan and Simpson (2001), anthropogenic contributions can be estimated based on fuel and electricity consumption, as well as the population and their related activities in the area of interest. For example, maximum values of 30 and 20 W m-2 of the sensible heat flux emission in the peak hours for vehicular and industrial contributions, respectively, were considered for the metropolitan area of São Paulo, Brazil, with more than 20 million inhabitants, and more than 7 million vehicles (Freitas et al., 2007). Such fluxes properly represent most of the urban heat island features for São Paulo, including the interaction between the UHI and the sea breeze. However, the fluxes must be adjusted on a case-by-case basis, and, therefore, urban structure and anthropogenic contributions are user-specified in the model namelist (Table 3) to limber model application for different urban areas. Additionally, the diurnal cycle of vehicle activities and other related features (pollutant emission, for example) are dependent on local time. Therefore, there is an input file describing local time as a function of the latitude and longitude of each grid point. Vehicular activity is defined in the model using a double normal distribution centered on two values of the time of rush hours, which are also user-definable (E. D. Freitas et al., 2005, 2007).

In this section, the coupling between the Joint UK Land Environment Simulator (JULES) surface–atmosphere interaction model (Best et al., 2011; Clark et al., 2011) and the BRAMS model is concisely described (for further details, readers are referred to Moreira et al., 2013). JULES contains the state-of-the-art numerical representation of surface processes and is able to simulate a number of soil–vegetation processes such as vegetation dynamics, photosynthesis, and plant respiration, and also transport of energy and mass in soils and plants, including a representation of urban elements. The coupling of JULES and BRAMS is fully two-way, with BRAMS providing atmospheric dynamics, thermodynamics, and chemical constituent information to JULES, which in turn responds with fluxes of horizontal momentum, water, energy, carbon, and other tracers exchanged between the atmosphere and the surface beneath. In JULES, the land surface is divided in sub-grid boxes, which can be occupied by a number of plant functional types (PFTs) and non-functional plant types (NPFTs). Up to five PFTs are allowed in each sub-grid box: broadleaf trees (BT), needleleaf trees (NT), C3 grass type (C3G), C4 grass type (C4G), and shrubs (Sh). A sub-grid box can also be occupied by up to four NPFTs: urban, inland water, soil, and ice. JULES adopts a tiled structure, in which the surface processes are calculated separately for each surface type. Its initialization requires land cover and soil type classifications, the normalized difference vegetative index (NDVI), sea surface temperature, carbon and soil moisture contents, and soil temperature.

Moreira et al. (2013) indicated that the application of JULES to simulations over South America implied a significant gain of skill compared to the original surface scheme in RAMS (LEAF3). As an example, Fig. 4 shows the model root-mean-square error (RMSE) of 2 m temperature, which was calculated using observations from ground stations distributed all over a large part of this continent. RMSE corresponds to the first 24 h forecast averaged over 30 runs in the wet (March, panel a) and dry seasons (September, panel b) of 2010. During the night, both surface schemes present similar skills, with LEAF3 being slightly better in the dry season. However, during daytime JULES notably improves model skills in both seasons. As a daily average, RMSE decreases by approximately 10 % with the latter surface scheme.

The shallow cumulus parameterization scheme in BRAMS is a mass flux type described in detail in Souza (1999). The cloud model follows the version of Albrecht et al. (1986) for a single-cloud formulation of the Arakawa and Schubert (1974) ensemble scheme. The shallow cumulus characteristic in the cloud model is obtained through an entraining function that gives more weight to the side entrainment as air parcels approach the cloud top. Therefore, a lifted air parcel from near the surface starts with a small entrainment of λ=10-6 m-1, and this value increases by an order of magnitude each time the parcel reaches a 10-fold height zf, which is the only adjustable parameter of the scheme. The entrainment rate is about 10-3 m-1 at the 2.1 km height for a zf of 0.7 km. The cloud top is reached when the total buoyancy of the parcel, integrated from the surface to the top, becomes zero. The mass-flux formulation is based on the heat engine framework proposed by Rennó and Ingersoll (1996). The derivation of the convective mass flux follows the rationale that the convective heat engine, which is driven by surface heat flux, forces the upward motion of air masses. The convective flux is then a result of the total forcing at the surface, namely the sum of the fluxes of sensible and latent heat, which are converted into kinetic energy according to the second law of thermodynamics. Once surface fluxes start forcing the heat engine, upward convecting air parcels might reach levels where water vapor saturation takes place. The triggering function follows the work of Wilde et al. (1985), which showed that moist parcels could give origin to shallow cumuli only when the entrainment zone, located on top of the mixing layer, is above the lifting-condensation-level zone.

This shallow convective scheme is suitable for studying the interaction between shallow convection and surface processes and its use in BRAMS improved the representation of the diurnal cycle of temperature and moisture over land.

The Grell and Deveny (2002, hereafter GD) deep convection scheme was included in BRAMS in 2002 and its implementation is described in S. R. Freitas et al. (2005). One of the reasons for the GD inclusion in BRAMS was the need for a mass flux scheme for consistent convective transport of tracers. GD expanded the original formulation based on Grell (1993) by including stochastic capability by permitting a series of different assumptions that are widely used in convective parameterizations. The GD scheme can use a very large number of ensemble members based on five different types of closure formulations, precipitation efficiency, and the ability of the source air parcels to overcome the convective inhibition energy.

Dos Santos el at. (2013) developed a method to generate a set of weights related to the closure members of the GD ensemble to optimize the combination of them. As an inverse problem of parameter estimation, the optimization problem for retrieving the weights applied a metaheuristic optimization method called the Firefly algorithm (FY, Yang, 2008). The method consists of minimizing an objective function computed with the quadratic difference between BRAMS precipitation forecasts and observation, a measure of the distance between the observational data and model results. Six different model simulations were performed to produce a five-member ensemble of precipitation forecasts, each one using a single closure option, and one of the runs was performed using the ensemble simple mean option. The single closure options used were Arakawa and Schubert (1974), moisture convergence (Krishnamurti et al., 1983), low-level Omega (Frank and Cohen, 1987), Kain and Fritsch (1992), and Grell (1993). The method proved able to produce an ensemble with improved statistical scores compared with the original ensemble mean calculation (Dos Santos et al., 2013; Santos, 2014). As an example, the categorical verification bias score computed for South America is depicted in Fig. 5. The mean of a set of 30 forecasts of 24 h accumulated precipitation for 120 h in advance of precipitation for January 2008 (panel a) and 2010 (panel b) was carried out using both the GD ensemble arithmetic mean (EN) and the ensemble mean using the FY method. The model setup included a grid with 25 km horizontal resolution covering South America and a 100 m vertical resolution in the first level; then, the vertical resolution varied telescopically with a ratio of 1.1 up to a maximum vertical resolution of 950 m, with the top of the model at approximately 19 km (a total of 40 vertical levels). As initial and boundary conditions, we used the CPTEC/INPE Atmospheric General Circulation Model (AGCM) analysis with T126L28 resolution, where T126 is the rhomboidal truncation at wave number 126 and L28 is the number of model vertical levels.

The vertical bars in Fig. 5 refer to a significance test from the bootstrap method (Hamill, 1999). These results indicate a reduction of bias at the low thresholds of precipitation, as well as an increase in the model skills for higher thresholds, in agreement with the increase in equitable threat score (not shown) for higher thresholds, both with statistical significance, which demonstrates that FY is a robust method for training the GD ensemble of closures.

In addition, the GD scheme in BRAMS contains an alternative option for the convective trigger function (CTF), which was originally developed by Jakob and Siebesma (2003) and implemented by Santos e Silva et al. (2012). In this formulation, the CTF is linked with the sensible and latent surface fluxes. Previous results, within both a global model (Betchold et al., 2004) and BRAMS (Santos e Silva et al., 2012), showed improvements in simulating the diurnal cycle of precipitation over continental areas, especially in tropical South America.

The Grell and Freitas (2014, hereafter GF) scheme is based on the stochastic approach originally implemented by GD, with several additional features. One new feature is scale-dependence formulations for high-resolution runs (or a gray zone for deep convection model configurations) and interaction with aerosols. The scale dependence was introduced by two approaches. One is based on spreading subsidence to neighboring grid points instead of in the same model convective column, as is usually done by classical convective parameterizations. The second approach applies methods devised by Arakawa et al. (2011). This work reformulated the eddy fluxes associated with the convective transports as a function of the updraft area fraction and the eddy fluxes given by a closure of a conventional convective parameterization. The idea is readily applied to the conventional parameterizations provided that a reliable formulation for the updraft area fraction is achieved. Because of its simplicity and its capability for an automatic smooth transition as the resolution is increased, Arakawa's approach is recommended to the BRAMS users.

A second new feature present in GF is an aerosol awareness capability through a CCN (cloud condensation nuclei number concentration) dependent autoconversion of cloud water to rain, as well as an aerosol dependent evaporation of cloud drops. However, this feature is still in the experimental stage, so caution when using it is advised.

Recently, the GF ensemble of closures has been extended to include a new closure inspired by ideas developed by Bechtold et al. (2008, 2014 – hereafter B2014). In the B2014 paper, the authors derive a diagnostic CAPE based closure where selective boundary layer timescales over land and water are applied. As a consequence, their convective parameterization improved its capability in the representation of non-equilibrium convection forced by boundary layer processes, with a more realistic phase of the associated diurnal cycle over land. In the GF scheme, 2015 version, a corresponding closure, although built on the cloud work function concept, was included. Additionally, GF, 2015 version, contains a variant scheme for shallow convection (non-precipitating) with three options for the closure of mass flux at the cloud base.

Several experiments with BRAMS, with the GF 2015 version, including Arakawa's approach (GF-A), using horizontal grid sizes of 5, 10, and 20 km, were carried out to evaluate the performance of the GF scheme as well as its behavior on different scales. For the 5 km model run we also described the performance of the scheme without applying any scale correction (GF-NS). Each experiment comprised 15 runs from 1 to 15 January for 36 h forecasts, all starting at 00:00 UTC; 24 h precipitation accumulation used for verification was taken from 12 to 36 h. Also, all experiments covered the same region and used the same initial and boundary conditions, which were taken from NCEP/USA Global Forecast System (GFS) analysis and forecast fields. Physical parameterizations included CARMA radiation, the JULES surface scheme, the Mellor–Yamada 2.5 turbulence scheme, and the single-moment bulk microphysics parameterization from Walko et al. (1995a). Model results are presented in Fig. 6. Decreasing the grid spacing from 20 to 5 km (panels a, b, and c), detailed precipitation structures show up, while the broad precipitation distribution is preserved with the domain-averaged precipitation, exhibiting deviation in a 10 % range (between 4.1 and 4.5 mm day-1). On the other hand, the precipitation produced by CP only (lower row, panels e, f, and g) presents a consistent decrease, becoming less significant, from 3.5 to 1.0 mm day-1, allowing the dynamics and cloud microphysics to be responsible for a much larger fraction of the total precipitation. Instead of a GF-A 5 km run, GF-NS (panel d) resulted in about 20 % larger domain average precipitation with a much smoother spatial distribution. In panel h is shown that, even on 5 km grid spacing, most of the precipitation (∼ 75 %) is generated by the convection scheme. These results demonstrate the ability of the GF-A scheme to produce a smooth transition across scales within the BRAMS modeling system.

Figure 7 introduces an exploratory study on the impacts of the B2014 closure (here called the “diurnal cycle” closure) on BRAMS results with respect to the diurnal cycle of precipitation over the Amazon Basin. The model configuration for this study comprised a grid with spacing of 27 km on the horizontal and 80 to 850 m on the vertical. The physical parameterizations and the initial and boundary conditions were the same as the preceding scale-dependence experiment, but GF applied the B2014 approach. Again, the model was set up to perform several runs resembling the operational mode, comprising 15 runs (from 1 to 15 February 2011) with 120 h forecast each.

Santos e Silva et al. (2009, 2012) discussed in detail the diurnal cycle of precipitation over the Amazon Basin using the TRMM rainfall product (Huffman et al., 2007) and observational data from an S band polarimetric radar (S-POL) and rain gauges obtained in a field experiment during the wet season of 1999. Their analysis indicated that a peak in rainfall is common late in the afternoon (between 17:00 and 21:00 UTC), in spite of variations existent associated with wind regimes. Figure 7 shows model results with and without the diurnal cycle closure; both panels depict area average precipitation from the GF scheme (mm h-1) as well as downwelling shortwave radiation (W m-2, DSWR). A sample of a 5-day forecast starting at 00:00 UTC, 1 February 2011, is presented in panel a. The simulated precipitation from the GF scheme not applying the diurnal cycle closure shows a premature peak, with both precipitation and DSWR closely in phase. The introduction of the B2014 closure causes a shift between the two curves, delaying the peak of precipitation by about 3 h, in better agreement with the observation. The diurnal cycle averaged over the 15 runs with 120 h forecasts each is presented in panel b, clearly showing the rainfall shift, which demonstrates the robustness of the B2014 closure. One potential drawback of this closure is the systematic reduction of the total amount of precipitation evidenced in panel b. Future work will focus on this issue.

An example of real-time rainfall forecast over South America with BRAMS using a different set of physical parameterizations is discussed as follows. The case is associated with a mid-latitude cold front approach together with tropical daytime convection over the northwestern part of the Amazonia Basin and a weak band of convection in the Inter-Tropical Convergence Zone (ITCZ) over the Atlantic Ocean. Figure 8 shows an estimate of the 24 h accumulated rainfall given by the TRMM product for the day of 12 October 2015 and depicts location and rainfall intensities of the cloud systems discussed above. This rainfall estimate is produced on a grid with 0.25∘ resolution. Model forecast was done on 5 km horizontal grid spacing with the vertical resolution varying from 50 m up to a maximum value of 850 m, with the top of the model at 19 km. The soil model was composed of seven layers distributed within the first 12 m of the soil depth. Again, GFS analysis and forecast fields were used for initial and boundary conditions, while initial soil moisture was supplied following Gevaerd and Freitas (2006), and the sea surface temperature was prescribed using data from Reynolds et al. (2002). The physical parameterizations included RRTMG shortwave and longwave radiation schemes, the GF 2015 version for deep and shallow convection with the diurnal cycle closure, Thompson single moment on cloud liquid water (no aerosol aware option) cloud microphysics, and the MYNN turbulence parameterization. The model run was completed on a CRAY XE-6 supercomputer using 2400 cores. This configuration took 1.6 h to complete a 24 h forecast with 1360×1480 on horizontal and 45 on vertical grid points, and 12 s for the time step. The simulation applied the hybrid time integration scheme with the RA time filter.

Figure 9 presents the 24 h accumulated rainfall model forecast for this day. The total (resolved plus from convection scheme) rainfall is shown in panel a. Visual comparison with TRMM rainfall (Fig. 8) shows that the model properly reproduces the main rainfall patterns over different parts of South America, despite the extreme amount of concentrated rainfall estimated by TRMM on the Amazon Basin (around 5∘ S and 65∘ W) being underestimated by the model. Similar model behavior is spotted in the Atlantic Ocean, close to the border between Brazil and Uruguay. However, in general, the model is able to capture consistently the rainfall intensity as well. Figure 9b shows the separated contribution of the cumulus convection scheme to the total rainfall (panel a). Noticeable is the fact that, on 5 km grid spacing, the scale awareness capability of the convection scheme allows the rainfall associated with the mid-latitude cold front to be almost entirely explicitly resolved. On the other hand, over tropical areas a significant part of the total rainfall is rather generated by the convection scheme, suggesting the existence of much smaller-scale rainfall systems, which is not explicitly captured at this model resolution.

The Coupled Chemistry-Aerosol-Tracer Transport model (Longo et al., 2013, hereafter CCATT) is an Eulerian transport model coupled with BRAMS and developed to simulate the transport, dispersion, chemical transformation, and removal processes of gases and aerosols for atmospheric composition and air pollution studies. CCATT computes the tracer transport in line with the simulation of the atmospheric state by BRAMS, using the same dynamical core, transport scheme, and physical parameterizations. The prognostic of the tracer mass mixing ratio includes the effects of sub-grid-scale turbulence in the planetary boundary layer and convective transports by shallow and deep moist convection, in addition to grid-scale advective transport. The model also includes gaseous/aqueous chemistry, scavenging and dry depositions, and aerosol sedimentation.

In a form of tendency, the general mass continuity equation for gas-phase tracers solved in the CCATT model is ∂s¯∂t=∂s¯∂tadv+∂s¯∂tPBLdiff+∂s¯∂tdeepconv+∂s¯∂tshallowconv+∂s¯∂tchem+W+R+Q, where s‾ is the grid box mean tracer mixing ratio, the term adv represents the 3-D resolved transport (advection by the mean wind) and the terms PBL diff, deep conv, and shallow conv stand for the sub-grid-scale turbulence in the planetary boundary layer (PBL) and deep and shallow convection, respectively. The chem term refers either to the simple passive tracers' lifetime (Freitas at al., 2009) or to the calculation of chemical loss and production (Longo et al., 2013). W is the term for wet removal applied only to aerosols, and R is the term for the dry deposition applied to both gases and aerosol particles. Finally, Q is the emission source term, which for biomass burning emissions also solves the plume rise mechanism associated with vegetation fires (Freitas et al., 2006, 2007, 2010).

In addition to CCATT-BRAMS code itself, the modeling system also includes three pre-processing software tools for user-defined chemical mechanisms (M-SPACK, Longo et al., 2013), aerosol and trace gas emissions fields (PREP-CHEM-SRC, Freitas et al., 2011), and the interpolation of initial and boundary conditions for meteorology and chemistry (BC-PREP) (see Fig. 10).

The choice of different chemistry mechanisms in CCATT-BRAMS is possible using a modified version of the SPACK pre-processing tool (Simplified Pre-processor for Atmospheric Chemical Kinetics, Damian-Iordache and Sandu, 1995; Djouad et al., 2002). The modified-SPACK (hereafter called M-SPACK) basically allows the passage of a list of species and chemical reactions from symbolic notation (text file) to a mathematical one (ODEs), automatically pre-processes chemical species aggregation, and creates Fortran 90 routines files directly compatible to be compiled within the main CCATT-BRAMS code. The M-SPACK output also feeds the codes of pre-processor tools PREP-CHEM-SRC and BC-PREP for emissions and the initial and boundary fields for the chemical species, respectively, in order to ensure consistency between the several input databases to be used in CCATT-BRAMS and the list of species treated in chemical mechanisms.

In principle, M-SPACK allows the use of any chemical mechanism in CCATT-BRAMS, though it requires building of the emissions interface. The current version of M-SPACK includes three widely used tropospheric chemistry mechanisms: RACM, the Regional Atmospheric Chemistry Mechanism (Stockwell et al., 1997), Carbon Bond (Yarwood et al., 2005), and RELACS, the Regional Lumped Atmospheric Chemical Scheme (Crassier et al., 2000), which consider, respectively, 77, 36, and 37 chemical species. Photolysis calculations are possible via LUTs of pre-calculated photolysis rates as well as through Fast-J (Wild et al., 2000; Brian and Prather, 2002) and Fast-TUV (Madronich, 1989; Tie et al., 2003) radiative codes. The latter approach provides online calculation of photolysis rates, including interaction of radiation with aerosols and clouds.

CCATT-BRAMS performance has been extensively evaluated for both urban and biomass burning areas (Freitas et al., 2009; Longo et al., 2010, 2013; Alonso et al., 2010; Bela et al., 2015). Figures 11 and 12 depict examples of model comparison results with mean daily values of carbon monoxide and ozone mixing ratio measured near the surface level in Porto Velho, Brazil, from 14 August to 8 October 2012. For this specific experiment, the model was configured to simulate smoke emission, transport, and its effects during the 2012 dry season in South America. The applied domain covered the whole of South America with a horizontal resolution of 25 km and 42 vertical levels. Atmospheric initial and boundary conditions were assimilated from analysis of the Brazilian Center for Weather Forecasting and Climate Studies global circulation model. The tropospheric chemistry mechanism used was RACM and biomass burning emissions for carbon monoxide and ozone precursors were estimated by the Brazilian Biomass Burning Emission Model (3BEM, Longo et al., 2010) in PREP-CHEM-SRC based on satellite remote sensing fire detections (Freitas et al., 2011).

BRAMS also has a simpler option for ozone forecasting suitable for urban areas. The Simple Photochemical Model (SPM) is available in the model together with the TEB scheme (E. D. Freitas et al., 2005). The model is composed of 15 reactions related to ozone formation and consumption. This small number of reactions was possible through the lamping of a large number of hydrocarbons, allowing a simplified way to deal with the photochemical process in the model, which is very convenient to be used in the operational mode. TEB-SPM considers industrial and vehicular emissions of carbon monoxide, volatile organic compounds (VOC), nitrogen oxides (NOx), sulfur dioxide (SO2), and particulate matter (PM2.5). In spite of its very simple formulation, the model has been used with relative success to simulate ozone concentrations in the São Paulo (E. D. Freitas et al., 2005) and Rio de Janeiro metropolitan areas in Brazil. Figure 13, adapted from Carvalho (2010), shows a comparison between model results and ozone observational data in two ground stations (Duque de Caxias and Jardim Primavera) of an automated network maintained by Rio de Janeiro's Environmental Agency (INEA). As one can see, the agreement is relatively high for a period over 7 days. For the simulations, the author used the Global Forecast System (GFS) analysis for the initial and boundary conditions. The model was set up with two nested grids of 16 and 4 km horizontal grid spacing, respectively, with 33 vertical sigma-z type levels. Both grids were centered at 22.80∘ S and 43.25∘ W. The coarser domain covered an area of 61 440 km2 (60×30 grid points), while the inner domain covered a 22 464 km2 area (54×26 grid points). The primary pollutant emissions were based on the inventories provided by INEA and considered both vehicular and industrial emissions for the five elements previously mentioned (CO, VOC, NOx, SO2, and PM).

This section introduces the capability of BRAMS composed with JULES in simulating CO2 fluxes associated with biogenic activities. Here we discuss an example of model simulation for September 2010 over the Amazon Basin. For this case, the BRAMS model was set with 20 km horizontal resolution covering the northern part of South America. The simulation was carried out for 45 days, starting on 15 August 2010 at 00:00 UTC, with the first 15 days discarded due to model spinup. The NCEP Global Forecast System analysis (http://rda.ucar.edu/datasets/ds083.2/), with 1∘ × 1∘ spatial resolution, provided initial and boundary conditions for the meteorological fields. The carbon data assimilation system, Carbon Tracker 2015 (Krol et al., 2005), with 3∘ × 2∘ horizontal resolution, provided the CO2 initial and boundary conditions. Biomass burning emissions of trace gases and aerosols were from 3BEM (Longo et al., 2010). The land use map, with 1 km spatial resolution, was provided by the USGS (United States Geological Survey), merged with a land cover map for the Brazilian legal Amazon region (Sestini et al., 2003). Figure 14 presents the gross primary productivity (GPP, panel a), plant respiration (PR, panel b), soil respiration (SR, panel c), and the net ecosystem exchange (NEE = PR + SR - GPP, panel d), all as a monthly average. September corresponds to the last month of the austral winter, with typically a very low amount of rainfall over a large part of Brazil. In this month, the ITCZ stays over positive latitudes, inducing rainfall only in the northwestern part of South America, with warm temperatures (maximum around 33 ∘C), low moisture, and clear skies. The abundance of photosynthetic active radiation and water availability in root zones of the tropical forest implies a large GPP over the region dominated by this land cover. As SR is mostly controlled by the soil humidity, the larger values are present in the region with higher rainfall amounts, which are in the northwestern part of the domain shown. At the same time, over areas dominated by Cerrado and Caatinga biomes, dry soil conditions dictate the response of the plants, with very low values of GPP and SR. However, the simulated NEE presents a more complex spatial distribution, with values oscillating from around zero and extreme around ±10 µmol C m-2 s-1, meaning CO2 in/out atmospheric fluxes (panel d).

BRAMS simulation of the diurnal cycle of CO2 in the low troposphere over the Amazon Basin is discussed as follows. Figure 15 shows 1-day simulation of the CO2 mixing ratio and the turbulent kinetic energy (TKE) in the low troposphere and DSWR at the surface. In this figure, TKE is used as a proxy for the depth of the atmospheric boundary layer, which evolves from a stable layer with less than 200 m depth during the nighttime and early morning towards a convective and well-mixed boundary layer with maximum heights of 1.2 to 1.5 km in the late afternoon. The results show a realistic nighttime near-surface accumulation of CO2 associated with the surface (soil and vegetation) respiration and a shallow stable boundary layer. After sunrise, with the increasing DSWR, the photosynthesis starts to dominate the net flux of CO2, which becomes more negative and subtracts this gas from the atmosphere. At the same time, the heating of the surface produces buoyant air parcels, which generates TKE deepening of the mixing layer. As a result, CO2 is mixed up and depleted inside of this layer, with its mixing ratio ending smaller than the one of the free atmosphere late in the afternoon.

The BRAMS tracer transport capability also incorporates emission, transport, dispersion, settling, and dry deposition of volcanic emissions, both for ash and a set of related gases. This capacity represents a critical step towards a numerical tool suitable not only for research, but also for an emergency, on-demand system for ash dispersion forecast after a volcanic eruption event, which is required for the safety of the air traffic around disturbed areas. The volcanic ash module follows closely the system described in Stuefer et al. (2013), and more details of its implementation in BRAMS are provided in Pavani (2014) and Pavani et al. (2016). The input needed to set up BRAMS for volcanic ash is produced using the PREP-CHEM-SRC (Freitas et al., 2011) emissions pre-processing tool, which contains a comprehensive database developed by Mastin et al. (2009). This database has information about 1535 volcanoes, including location (geographical position and height above sea level of the vent) and a set of historical parameters (e.g. initial plume height, mass eruption rate, volume rate, duration of eruption, and size distribution of the ash particle), which can be used as a first guess for a potentially recurring volcanic eruption. However, by default, whenever available, observed real-time information overwrites the historical ones. In BRAMS simulations, a vertical profile of the ash emission distribution is defined by a linear detraining of 25 % of the total ash mass below the injection height and 75 % around it, obeying a parabola shape. Pavani et al. (2016) adjusted an exponential curve between the rate of ash mass produced during the eruption and the injection height, which is expressed as follows: H=0.34M0.24, where H is the plume height in km (height above the vent) and M is the emission rate in kg s-1. This fitting formula is an additional method to make a first guess of the erupted mass of ash when the injection height is known.

The model functionality for volcanic ash dispersion has been applied to real cases. One example is the eruption of the Puyehue volcano in Chile, which occurred around 20:15 UTC on 4 June 2011, expelling a huge mass of ash and gases up to 13 km in height above sea level. This eruptive event caused the closure of numerous airports for many days and transport disruption in several countries in South America, South Africa, and even Australia and New Zealand. Additionally, ash scavenging caused harm to agriculture and livestock, besides other economical and public health related issues. Costa et al. (2012) described the development and application of a remote sensing technique for traces of ash retrieval based on METEOSAT-8 satellite data.

The BRAMS simulation to study the transport of the Puyehue volcano ash was carried out for 40 days starting on 4 June 2011, 00:00 UTC, with 30 km horizontal resolution and a vertical resolution starting at 100 m at the surface, stretching with a ratio of 1.1 up to 500 m at the model top. Figure 16 shows the location of ash as determined by this technique on 6 June 2011, 15:00 UTC, approximately 44 h after the first eruption event. The eruption introduced material into the jet stream region which was rapidly transported eastward following the Rossby wave circulation. BRAMS results for this case study showed significant improvement with the use of the monotonic advection scheme described in Sect. “Monotonic scheme for advection of scalars”, since monotonicity is required to properly model the long distance transport of tracers associated with sharp, small-scale emission sources within low-resolution atmospheric models. Figure 17 depicts the regional distribution of the ashes, as an example of the model performance in simulating the long-range transport of volcanic emissions. Panel a shows the simulated vertically integrated total mass of ash (mg m-2) for the same time of the remote sensing imagery (Fig. 16). Panel b shows the ash total mass concentration (µg m-3) at approximately 9500 m above the surface. At the beginning of the eruption, the ash was transported eastward for about 20∘, and then assumed an undulating shape associated with the Rossby waves. The ash layer at 9500 m is constituted primarily of small-sized particles, since the larger and heavier ones quickly fall vertically due to the gravitational force. The higher sensitivity of the ash retrieval in the upper levels explains the better agreement between the ash distribution presented in this panel and the traces of ash retrieved by remote sensing (Fig. 16). The wider ash distribution close to the volcanic vent in panel a is associated mainly with the vertical settling of the large, heavy ash particles that end by getting different wind circulations and/or are quickly deposited over land. A more comprehensive analysis of this case study is discussed in Pavani et al. (2016).

The Stochastic Time-Inverted Lagrangian Transport model (STILT, Lin et al., 2003) is a Lagrangian model framework coupled with surface emission models, and has been used to identify sources and their influence on receptors in studies with a multitude of scales and chemical components (see Gerbig et al., 2003; Miller et al., 2008, 2013; Xiang et al., 2013; McKain et al., 2015). The core component of STILT is a Lagrangian particle dispersion model that has two key features that allow for a realistic representation of dispersion: (1) STILT accounts for sub-grid-scale transport and dispersion by incorporating an stochastic component associated with small-scale turbulence (Lin et al., 2003); (2) STILT also accounts for vertical transport due to parameterized convective clouds (Nehrkorn et al., 2010). However, in order to take full advantage of BRAMS turbulent and convective models, additional turbulence and convection related quantities are included in BRAMS output so that they can be directly used by STILT.

Following Lin et al. (2003), in STILT each wind component ui can be decomposed following a Markov assumption, i.e. the grid volume average component u′i and a turbulent component u′i. The turbulent component is modeled after Hanna (1982), who defines the autocorrelation coefficient in terms of the Lagrangian timescale TLi and the standard deviation of wind σui in the mixing layer: ui′t+Δt=αiΔtui′t+N0,σuit1-σiΔt2,αiΔt=exp-ΔtTLi, where t is the previous time, Δt is the time step, and N is a random number following the normal distribution with mean 0 and standard deviation given by σui. For consistency with the turbulence scheme, the standard deviation is computed following Nakanishi and Niino (2004). The Lagrangian timescale is determined following the parameterization by Hanna (1982), which also depends on the boundary layer depth. Hanna (1982) parameterization of the boundary layer depth depends on the reciprocal of the vertical component of Coriolis vorticity, which would cause singularities at the Equator. Therefore, we implemented an alternative parameterization by Vogelezang and Holtslag (1996).

When BRAMS simulations are carried out using the Grell and Dévényi (2002) cumulus parameterizations, all mass fluxes associated with updrafts and downdrafts (entrainment, detrainment, and vertical motion) are also saved to the output, and can be used to assign both the probability of any particle being in the environment or in the cloud (either at the updraft or downdraft), as well as the vertical displacement of particles in case they are in the updrafts or downdrafts, using the same method described by Nehrkorn et al. (2010). Besides, the inclusion of mass flux and turbulence related variables in the output also allows a seamless integration with different Lagrangian particle dispersion models.

BRAMS simulated fields can readily be applied as input data to a 3-D air parcel kinematic trajectory model described in Freitas et al. (1996, 2000). Forward and backward time integrations are allowed using a second order in time accurate scheme. The trajectories are computed using the same map projection and the vertical coordinate of BRAMS and also include a sub-grid-scale vertical velocity enhancement associated with sub-grid-scale convection not explicitly solved by model dynamics.

A digital filter for model initialization has been implemented in BRAMS and demonstrated the ability to reduce high-order imbalances and inconsistencies among model variables, with the potential to improve deterministic forecasts.

A new feature present in BRAMS is the possibility of the model output being produced in GrADS (http://iges.org/grads) format during the runtime, simultaneously with the model integration. This feature is especially important for operational centers by allowing faster generation of operational products.

Spread Fire (SFIRE) is a semi-empirical fire propagation model developed by Coen (2005), Clark et al. (2004), and Mandel et al. (2009, 2011) that was coupled to the BRAMS model and is currently under evaluation. SFIRE simulates a fire propagation based on a spread rate S=Sx,y,t in an orthogonal direction to the fire boundary and expressed as a function of the wind v=vx,y,z,t and terrain gradient ∇z. The model provides the sensible and latent heat fluxes associated with the fire propagation (the second terms of the RHS of Eqs. 23 and 24, respectively), allowing feedbacks between the combustion processes and the surrounding atmosphere. The total sensible QH and QE latent heat fluxes are given by QH=-cpρaT∗u∗+FT-Ft+ΔtΔt11+Mfwh,QE=-χ∗ρau∗+FT-Ft+ΔtΔtMf+0.561+MfwL. Here the fluxes depend on properties of forestry fuel models, following Anderson (1982) categories, and on an exponential decay function of total fuel fraction, F (see Table 4 for a detailed description of the symbols). The fuel fraction decreases exponentially from the initial ignition time ti (Albini, 1994) and is given by Fx,y,t=1area∫∫(x,y)∈Ω(t)e-0,8514t-tix,ywx,ydxdy,1otherwise. An interface code was built to implement a link between SFIRE and BRAMS, which includes new modules for memory allocation, initialization, and a new namelist (sfire.in). Currently, the fire spread model runs only in a serial mode inside of a BRAMS parallel simulation. Full parallelization of the SFIRE model will be available in future BRAMS versions.

The user needs to produce fuel model classes map and topography defined on the refined surface meshes used by SFIRE; to do so, the user must download any necessary high-resolution fields (topography raster and FNNL fuel models; Anderson, 1982) and convert them into a geographic information system (GIS) for ASCII (American Standard Code for Information Interchange) format, through a Euclidean allocation interpolation. Instead the user can use topography from BRAMS, although it is highly smoothed for the needs of SFIRE, and the code cannot benefit for more accurate fire spread computations, because it required a high-resolution grid. These high-resolution data are interpolated and assimilated by BRAMS-SFIRE in fire mesh simulations.

The atmospheric data available to BRAMS are limited to around 111 km resolution and should be simulated in downscaling grids, each with a 4-to-5 refinement ratio, and can incorporate weather sounding data. The BRAMS model was simulated on a 3-D grid covering the Earth's surface, and only the downscaling refined atmospheric domain can be activated with the SFIRE model. BRAMS-SFIRE was applied to the region of Alentejo in Portugal, but can be applied to any other region of the world. The coupled model has an input file named “namelist.fire” where the user is able to introduce the properties of fuel models of the region of interest (Menezes, 2015).

The average sensible and latent heat fluxes released in the time interval t,t+Δt, Eqs. (23) and (24), from SFIRE are passed into the atmospheric model through fluxes coming from boundary conditions and mixed in the boundary layer by the PBL scheme.

The BRAMS-SFIRE simulations were performed using the downscaling procedure (one-way interaction), which started from a model grid of 64 km resolution (with the model domain covering Europe). The data from this simulation are then applied to feed another model run with a grid of 16 km resolution (covering continental Portugal), which in turn feed another grid of 4 km resolution (covering the Alentejo), which feed another grid of 1 km resolution (which covers the area under study). Finally, the 200 m grid resolution simulation (in the area of forest fire) applied the SFIRE model with atmospheric fields provided by the 1 km resolution grid model run.

The BRAMS physical parameterizations were configured with silhouette orography to a topography scheme, Klemp–Wilhelmson, to lateral boundary conditions, CARMA to shortwave and longwave radiation schemes, with a 900 s frequency update of the radiation trend, microphysics complexity level 3 (Flatau et al., 1989), Grell 3-D formulation convective parameters with a convection 900 s frequency, and Grell–Deveny parameters of shallow cumulus with a 1200 s frequency. For 1 km and 200 m grid resolutions, cumulus parameterization was not used. The JULES surface scheme (Moreira et al., 2013) was used for downscaling until 4 km resolution, and LEAF (Walko et al., 2000) was used in the 1 km and 200 m resolution grids. The turbulent diffusion coefficient parameter of Mellor and Yamada (Mellor and Yamada, 1982) was used in all grids until 1 km resolution, and an isotropic deformation was used for the 200 m resolution grid.

The simulations were carried out with the non-hydrostatic equations on a vertical grid with 55 levels. SFIRE was configured with 200 m horizontal grid resolution and by updating the fuel moisture calculation every 30 s, with a reaction velocity of 7 m s-1, with a 15 km radius prescription fire, and with a fire initiation time of 180 s after the start of the atmospheric simulation.

One of the results from the BRAMS-SFIRE simulation showed that, over the three regions, of flat land and low hills, the propagation of the fire line originated sensible heat fluxes of approximately 28 kWm-2. During its spread over fuel models 1 and 2, the fire burned them quickly, compared to fuel model 4, which degraded fuel and released fluxes of about 2.5 kWm-2 during combustion. Fuel models 8 and 9 combustion-liberated fluxes of between 1.4 and 1.6 kWm-2, and over fuel model 9 liberated fluxes on the order of 1.2 kWm-2, which glowed until its extinction at approximately 0.75 kWm-2. The fire spread was influenced by the topography gradient, following dispersion over valleys or down the mountain (Ossa mountain range, Fig. 23) or simply propagating into plain zones. In all three regions, propagation occurred in an elliptical pattern. In the Ossa mountain range region, the wind is anabatic with intensity 4.5 ms-1, and changes its pattern with fire outbreak, becoming disordered with vortices on the fire which increased the intensity of the wind to 7.5 ms-1; this pattern extends to the entire region as the fire develops and the fire line spreads (Fig. 23).