2.1 Gridded emission inventories

Emission inventories of greenhouse gases and air pollutants are usually provided as annual 2D gridded fields per source category (e.g., traffic, industry, residential, agriculture) (Kuenen et al., 2014; Janssens-Maenhout et al., 2017; Crippa et al., 2018). Before incorporating these emissions into a 3D atmospheric transport model, the 2D fields have to be convolved with temporal profiles of diurnal, day-of-week, and seasonal variability as well as with vertical emission profiles (Denier van der Gon et al., 2011; Pouliot et al., 2012; Kuenen et al., 2014). Current approaches for the definition of temporal and vertical profiles are often highly simplified (Matthias et al., 2018), partly due to a lack of detailed information but also due to the limited flexibility offered by the offline approach. Our current implementation of the module supports three different families of emission inventories: the inventories of the Netherlands Organization for Applied Scientific Research (TNO), the EDGAR inventories of the European Joint Research Centre (JRC), and the Swiss national emission inventories generated by Meteotest Inc., Switzerland. An extension to similar gridded inventories should be straightforward.

Although the inventories share a similar logic, there are distinct differences that had to be accounted for: in Europe, the two most prominent source classifications are the Standardized Nomenclature for Air Pollutants (SNAP; Centre on Emission Inventories and Projections, 2018) and the gridded Nomenclature for Reporting (GNFR; United Nations Economic Commission for Europe, 2015; Schindlbacher et al., 2016). SNAP is an old standard introduced in the context of the Convention on Long-Range Transboundary Air Pollution (CLRTAP), which had been used for many years in inventories produced by the European Monitoring and Evaluation Programme (EMEP) and by TNO. The latest inventories, however, follow the EMEP/CORINAIR NFR, which was introduced to harmonize the source classification with the one used in National Inventory Reports to the United Nations Framework Convention for Climate Change (UNFCCC EEA, 2000). For gridded inventories, the large number of NFR categories are usually lumped into 13 individual classes (A–M). This reduced set of categories is commonly referred to as gridded NFR or GNFR. EDGAR inventories are reported as individual or lumped NFR categories, which are similar but not identical to the GNFR standard. A table for unambiguous mapping between individual NFR, GNFR and SNAP categories is available at the United Nations Economic Commission for Europe (2003).

Another difference is the map projection. EDGAR and TNO inventories are reported on a regular latitude–longitude grid and the Swiss Meteotest inventories on a Cartesian grid in the Swiss CH1903 oblique Mercator projection. Furthermore, TNO inventories differentiate between point and area sources, which is not done in the other inventories. Point sources as reported to the European Pollutant Release and Transfer Register E-PRTR (https://prtr.eea.europa.eu, last access: 15 May 2020) are provided in the TNO inventories at their exact location, which has the advantage that in high-resolution simulations the point sources can be assigned accurately to the respective model grid cell. In addition, this allows applying separate vertical profiles to point and area sources as shown in Sect. 3.3.2.

2.2 Mapping to COSMO grid

All emission data need to be mapped onto the simulation grid, which in the case of COSMO is a rotated latitude–longitude grid. Such mapping is straightforward for point sources for which the emissions are added to the COSMO grid cells that contain the sources. For area sources this is less trivial, since simple interpolation is not mass-conservative, whereas conservative nearest-neighbor methods may lead to undesired stripes or other discontinuities. In order to avoid such issues and to accurately conserve mass, we determine the relative fraction of the overlap with each inventory grid cell for each COSMO grid cell. The emissions for all source categories s at each COSMO grid cell i are then computed by

where $E_{i,s}^{\mathrm{COSMO}}$ and $E_{j,s}^{\mathrm{inventory}}$ are the emissions (in mass∕cell) at the ith and jth COSMO and inventory grid cell, respectively, N is the number of grid cells in the inventory, and f_i,j is the dimensionless fraction of the source cell j contributing to destination cell i. The fraction is determined by computing the area of the intersection between each COSMO and inventory grid cell divided by the total area of the inventory cell using the equal-area Mollweide projection to conserve mass.

Our tool does not redistribute the low-resolution inventory data onto the high-resolution COSMO grid using additional information, such as land–sea masks and country boundaries, which can be used for improving the spatial allocation of area sources. This feature could be implemented in a future version similar to the implementation in the CHIMERE-2017 model (Mailler et al., 2017).

The generated netCDF file contains 2D gridded fields E_i,s directly on the COSMO grid. This file also contains a corresponding 2D country mask, which is an integer field with each grid cell being assigned the number of the country that has the largest fractional area. The unit of the emissions $E_{i,s}^{\mathrm{COSMO}}$ depends on the actual model and is converted accordingly. For example, COSMO-GHG expects kilograms per square meter per second, whereas emissions in COSMO-ART are in kilograms per hour per cell.

2.3 Temporal and vertical profiles

The emission of a tracer X at time t is calculated as the sum over the emissions per source category scaled with source-specific temporal scaling factors. A tracer is defined here as a quantity that is transported in the model, which can be a single trace gas or aerosol component, a composition of several species, or an idealized tracer. The temporal scaling factor for tracer X, source category s, and time t is generally given by

where $w_{X,s,h}$, $w_{X,s,d}$, and $w_{X,s,m}$ are diurnal, day-of-week, and seasonal scaling factors, respectively. The three scaling functions are dimensionless and have a mean value of 1, such that the mean value of all scaling factors applied to a full year of data is 1 (or very close to 1). The step functions h(t), d(t), and m(t) are the hour of the day, the day of the week, and the month of the year corresponding to the continuous time t, respectively. The emission of X at time t is thus

where E_X,s is the annual mean emission flux of X of source category s (which is the basic field usually provided by an inventory) and n is the total number of source categories. This formula is applicable to an emission from a single grid cell or to a complete 2D emission field; i.e., E_X and E_X,s may be 2D fields. The functions may further depend on the country of the source. In that case, a further summation over countries is needed in combination with country masks. Our Python package and online emission module support country-specific time functions.

The generated netCDF files contain time functions of diurnal, day-of-week, and seasonal variations per tracer and source category. These scalings are provided in three separate files by default. However, it is also possible to provide only one file with hour-of-year scaling factors. The temporal profile variables are arrays with the two dimensions time (e.g., 24 different hourofday in the case of a diurnal profile) and country. If no country-specific information is available or desired, a uniform country mask with a single value for the whole model domain needs to be generated. Vertical profiles, in contrast, are 1D arrays with level as the only dimension, since they are not expected to vary with country.

With this approach, real trace gases such as CO₂ can be simulated but so can idealized tracers representing only a subset of sources, for example a tracer representing only traffic CO₂ emissions, by only summing over a subset of source categories in Eq. (3).

Emissions do not only occur at the surface but should be treated in 3D (Bieser et al., 2011b; Brunner et al., 2019). This is particularly true for elevated emissions from power plants or air traffic. Idealized vertical scaling functions v_s are available for anthropogenic emissions, which distribute the emissions from a source of category s over a discrete set of geometric vertical layers (altitude relative to ground). The scaling factors add up to 1 when summed over all vertical layers. Examples are given in the Supplement. The emission of the simulated tracer X at time t and in vertical layer k is then given as

The file generated by the Python tool contains vertical profiles per tracer and source category. The number of levels and their heights above surface can be set independently of the vertical structure of the COSMO grid. The vertical profiles do not depend on time t in the current implementation of the module. This could be implemented in the future, for example to account for a meteorology-dependent plume rise of emissions from power plants.

2.4 Speciation

The chemical compounds simulated in COSMO-ART include species for which inventories provide direct emission strengths (e.g., SO₂ or NH₃). However, for other species, the inventories only provide aggregated information for a family of compounds. This is the case for NO_x (sum of NO and NO₂), NMVOCs, and particulate matter with a diameter of less than 2.5 µm (PM_2.5) and 10 µm (PM₁₀) (sum of various organic and inorganic aerosol compounds).

Therefore, to compute the emission of an individual compound simulated in the model, chemical speciation factors have to be applied to the total mass of the family. These speciation factors are specific for different source categories, since, for example, the composition of NMVOCs, PM, and NO_x emissions is different for traffic and residential heating. Furthermore, the speciation factors depend on the specific chemical mechanism applied in the model, which determines the mapping between real and model-simulated species. Starting from Eq. (4), considering a simulated tracer X whose emissions are reported under a chemical compound $\stackrel{\mathrm{̃}}{X}$ with a speciation factor $f_{\stackrel{\mathrm{̃}}{X}\to X}$ (dimensionless), the emission of X at vertical level k is given by

3.1 Basic framework for online emission processing in COSMO

The main philosophy is to read in all input data required for the online emission module only once at the start of the simulation. These data include annual mean sector-specific 2D emission fields E_X,s as well as the temporal, vertical, and speciation profiles. During the simulation, these profiles are applied online to update the hourly emissions for each species according to Eqs. (2)–(5).

Table 2Members of the GHGCTL and OAECTL namelist group required in INPUT_GHG for COSMO-GHG and INPUT_OAE for COSMO-ART, respectively, for setting up the online emission module.

Download Print Version | Download XLSX

Table 3Namelist members of the TRACER group in INPUT_GHG and INPUT_OAE. Some namelist members are only implemented in COSMO-GHG or COSMO-ART.

Download Print Version | Download XLSX

In order to implement new tracers flexibly, TRACER groups were added to the INPUT_GHG namelist file. A subset of the possible entries (members) of the TRACER group is presented in Table 3 and an example namelist file is given in the Supplement. For the online emission module, the definition of the TRACER group was extended with the following parameters:

1. a new possible value for the switch itype_emiss, which needs to be set to 2 to activate online emissions for this tracer

2. a list tag ycatl listing the categories s considered as sources of the tracer

3. a list tag ytpl listing the temporal profile used for each element of ycatl

4. a list tag yvpl listing the vertical profile used for each element of ycatl

5. a list tag contribl used for chemical speciation, which lists the contribution of categories s to the total emitted mass of the tracer (only COSMO-ART).

Note that the comma-separated lists ycatl, ytpl, yvpl, and contribl need to have the same length. In contrast to COSMO-GHG, the definition of tracers is fixed (i.e., hard-coded) in COSMO-ART, and, therefore, no namelist file INPUT_GHG exists. In order to enable the same functionality as in COSMO-GHG, a new namelist file INPUT_OAE was introduced, where for each emitted tracer a corresponding TRACER group has to be defined (OAE: online anthropogenic emissions). In order to enable or disable the use of online emissions in COSMO-ART, a new Boolean switch lemiss_online was implemented in the namelist file INPUT_ART, which needs to be set to .TRUE. to activate the online emission module. For this reason, the namelist switch itype_emiss does not exist in COSMO-ART. The same applies for itype_tscale, since temporal scaling is applied just for hour-of-day, day-of-week, and month-of-year profiles.

At the start of a simulation, the online emission module reads in the emission fields and the temporal and vertical profiles from the netCDF files generated by the Python tool, which have been described in Sect. 2. Full paths of the files have to be specified in the namelist file INPUT_GHG or INPUT_OAE; see Table 2. Note that variable names within the netCDF files have to be identical with those listed in ycatl, ytpl and yvpl, as shown in Table 4.

Figure 2Flow chart for COSMO-GHG run-time processes both for the offline and the online approach. Blue rectangles are code parts. The simulation start point is represented by an oval.

Download

Table 4Overview of netCDF files and corresponding namelist tags in the TRACER group in INPUT_GHG or INPUT_OAE. The netCDF variable names must be identical to those listed in the namelist members (right column).

Download Print Version | Download XLSX

Speciation profiles for COSMO-ART are not provided in a separate netCDF file but are included in the namelist file INPUT_OAE through the list tag contribl of the TRACER group.

3.2 Modifications to the COSMO-GHG and COSMO-ART codes

In order to enable the computation of online emissions as an alternative to the default offline reading of emission files, the greenhouse gas module of COSMO-GHG (m_online_emissions.f90) was extended with new structures to store the information on temporal and vertical profiles and the annual mean sectorial emission grids. The arrays for temporal profiles and the emission grids are dynamically allocated as they depend on the tracer type. The information from the netCDF files is read at the beginning of the simulation. The vertical scaling factors, which are defined for layers above ground of fixed vertical extent (see Supplement), need to be translated into scaling factors for the vertical layers of COSMO. Since COSMO uses a spatially fixed grid with geometric hybrid vertical layers with thicknesses varying with the underlying topography, the vertical scaling factors are translated into dynamically allocated 3D arrays.

Before the implementation of the online emission module in COSMO-GHG, emissions were read-in every hour from a file and assigned to an emission field associated with each tracer. In COSMO-GHG, fields that are conceptually attached to a prognostic variable of the model (e.g., an emission field attached to a tracer) are called associated fields. The definition, memory management, and I/O aspects of associated fields are handled in src_associated_fields.f90. There, besides the already existing functionality of reading in files, a mechanism was added to compute the emissions based on the dynamically allocated arrays for the emission grid and temporal and vertical scaling factors. Afterwards, no matter if online or offline emissions are used, the emission fields are updated at the beginning of each hourly interval. The updates account for the new temporal scaling for the current hour and the 3D vertical scaling factors calculated at the beginning of the simulation by applying Eq. (4). The emissions are then kept constant during the current hour or linearly interpolated in time between the current and the next hour (if iemiss_interp is set to 1). The latter requires the computation (or reading) of the emission field at the following full hour. Finally, the emission field is added to the tracer tendencies at each model time step. All relevant computations in the GHG module use OpenACC directives in a similar manner as in other parts of the main COSMO code. A schematic of all relevant processes for offline and online emissions in COSMO-GHG during model run time is shown in Fig. 2.

Even though the concepts are very similar, the realization of the online emission module in COSMO-ART differed significantly from the implementation in COSMO-GHG. The emission module in ART (art_emiss_prescribed.f90) comes on top of the tracer module in COSMO and interacts closely with the tracer fields in other ART modules such as art_mademod.f90 for aerosol processes. It was therefore desirable to keep most of the functionality of the standard emission modules to ensure the usability of both online and offline emission versions. The online emission module in COSMO-ART works as follows, focusing on the differences to the COSMO-GHG implementation: During initialization of the model, if the namelist parameter lemiss_online is set to .TRUE., the required files (emissions, temporal and vertical scaling factors) are read-in as described in Sect. 3.1. For every tracer with online emissions, the gridded emissions, the scaling factors (temporal and vertical), speciation values are collected in a data structure that associates this information with the location of that tracer in the art_species structure. Then, at hourly intervals during the run, the emissions are calculated and written into the corresponding array in art_species. Because this replaces the assignment of emissions read-in from external files to the same array as done in the standard offline version, no further changes to other modules are required. It is even possible to mix on- and offline emissions in a simulation run, such that emissions of certain tracers are read-in, while others are calculated.

3.3 Practical examples

4.1 COSMO-GHG

Using the COSMO-GHG model with online and offline setup, respectively, a simulation with a single CO₂ tracer for a domain covering the Alpine region at $\mathrm{0.01}{}^{\circ }\times \mathrm{0.01}{}^{\circ }$ (≈1.1 km) horizontal resolution and with 60 vertical levels was conducted. The simulation extended over a period of 1 week from 1 to 8 January 2019. The boundary conditions for meteorological fields were taken from COSMO-7 reanalyses provided by MeteoSwiss. The CO₂ tracer field was initialized to 0.

The model setup closely followed the settings of the operational high-resolution COSMO-1 forecast of MeteoSwiss. The simulations were performed on 60 hybrid CPU–GPU nodes of the Piz Daint supercomputer of the Swiss National Supercomputing Centre (CSCS). Each node included an Intel Xeon E5-2690 v3 12-core CPU and an NVIDIA Tesla P100 GPU.

The yearly gridded anthropogenic emissions of CO₂ were taken from the TNO-GHGco (Super et al., 2020) inventory for 2015 available at $\mathrm{0.1}{}^{\circ }\times \mathrm{0.05}{}^{\circ }$ resolution. They were split into 12 GNFR categories. The F category for road transport was further divided in three subcategories depending on the type of vehicles. Over Switzerland, the TNO inventory was replaced by a Swiss CO₂ inventory at 500 m×500 m resolution, which was created by the company MeteoTest in the framework of the CarboCount-CH project (Liu et al., 2017). Due to a different source categorization used in the Swiss inventory, those categories were mapped to the GNFR categories B, C, F, J, and L. The temporal profiles applied to the emissions are described in the Supplement. No vertical profiles were applied in this example, but all emission were released at the surface. The namelist definition of the tracer in this simulation is described in the Supplement.

An example of the distribution of the simulated CO₂ tracer representing all anthropogenic emissions in the domain and of the differences between online and offline is shown in Fig. 3. The figure shows instantaneous near-surface CO₂ at the end of the 7 d simulation period.

Figure 3Surface-layer concentrations for the online simulation (a) and relative differences between online and offline simulations (b) of the CO₂ tracer corresponding to all anthropogenic emissions after 7 d in the COSMO-GHG test case.

The differences between online and offline are negligible (around 10⁻⁶ ppm) almost everywhere except for a plume in southern Germany. This difference can be explained by a point source in the TNO inventory, which is located exactly on the border of a COSMO grid cell. When generating online and offline emissions, two separate gridded maps of emissions need to be produced.¹ Because of floating point truncation errors, this point source was attributed to two adjacent grid cells when processed with the online and offline approach. The spatial shift of this plume leads to differences of around ±0.15 ppm but the spatial mean remains almost constant (differs by less than 10⁻⁷ ppm).

In terms of computation time the two COSMO-GHG runs were almost identical, as seen in Table 5. Disk usage, on the other hand, was dramatically reduced when using online emissions (about 3 % of the offline case). This benefit would have been even larger for a longer simulation period. Furthermore, the time consumption for generating the online input files is more than halved compared to the hourly offline files, as shown in Table 6.

4.2 COSMO-ART

The online and offline emission approaches of COSMO-ART were compared by performing a test simulation over Europe with a horizontal resolution of $\mathrm{0.12}{}^{\circ }\times \mathrm{0.12}{}^{\circ }$ (≈13 km) and 60 vertical levels. The simulation was driven by meteorological fields from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecast System (IFS) model. The initial and boundary conditions for the chemical species were taken from the global MOZART-4 model (Emmons et al., 2010). A standard configuration with the RADMK chemical mechanism, Volatility Basis Set (VBS) for organic aerosols, and ISORROPIA-II scheme for inorganic aerosols was selected as described in Athanasopoulou et al. (2013). RADMK is an extended version of the RADM2 chemistry scheme (Stockwell et al., 1990) with an improved representation of isoprene chemistry (Vogel et al., 2009). The configuration for the meteorology closely followed the setup of the operational European COSMO-7 forecasts of MeteoSwiss. The 24 h test simulation was started at 26 June 2015, 00:00 UTC, and ended at 27 June 2015, 00:00 UTC. It was conducted on 16 CPU nodes of Piz Daint. Each node consisted of two Intel Xeon E5-2695 v4 processors providing a total of 36 cores per node. The simulation thus used a total of 576 cores.

The European anthropogenic emission inventory CAMS-REG-AP_v2_2 (Granier et al., 2019) generated by TNO was combined with the Swiss national emission inventory for reactive gases and aerosols generated by Meteotest Inc. Both inventories were based on the GNFR source classification. The two inventories were merged using a country mask for Switzerland, i.e., following the first approach described in Sect. 3.3.1.

The COSMO-ART simulations with online and offline emissions have almost identical outputs for the same test simulation as shown in Fig. 4 for surface-layer concentrations of the gas phase species SO₂ and ethane and for particulate sulfate (SO₄, variable VSO4J in COSMO-ART). The differences for SO₂ and ethane are in the order of numerical noise. For SO₄ the differences are larger but still several orders of magnitude smaller than the absolute concentrations. Due to the complexity and non-linearity of aerosol chemistry within COSMO-ART, small differences may eventually build up in the course of the simulation. While maximum relative differences of SO₄ are up to 10 % for a few grid points, the spatial mean of these differences is small with 0.044 %. Further statistical values for SO₄ and other variables are provided in the Supplement.

Similar to COSMO-GHG, the computation time was quite comparable for the two versions, as seen in Table 5. The reduction in time spent on I/O in the online version is thus largely compensated for by the increase in computation time. Disk usage for the online emission version was only about 1 % of the usage for the offline version. Again, the benefit in disk usage would grow proportionally with increasing length of the simulation period.

Figure 4Surface-layer concentrations for the online and offline simulations and relative and absolute differences between online and offline simulations in the COSMO-ART test case for surface layer fields of SO₂ (top), SO₄ (middle), and ethane (bottom) after 24 h.

Table 5Simulation time and input data size for online and offline COSMO-GHG and COSMO-ART test cases.

Download Print Version | Download XLSX

Table 6Benchmark for pre-processing online and offline emissions for the COSMO-GHG test case using the emiproc tool. Altogether, 19 categories from the emission inventories were processed (TNO: 14; Swiss: 5) as input for the CO₂ tracer. Pre-processing was performed on a local Linux cluster, using 14 threads in parallel. Processing times for generating the mapping and country mask files are excluded. Results are shown for different dataset lengths.

Download Print Version | Download XLSX