We first provide an overview of the gridding methodology, with further details provided in the following subsections. The gridding methodology for historical anthropogenic and all future emissions is summarized in Fig. 1. The fundamental underlying data used here are emissions by country and sector. This provides both total emission trends over time as well as changes in the sectoral composition of each emission species. As discussed in Hoesly et al. (2018), emissions in the preindustrial period were generally dominated by biofuel use in the residential sector. As industrialization proceeded emissions from the industrial, energy transformation, and transportation sectors became increasingly important. Emissions are translated to a spatial grid for each country and aggregate gridding sector (see Table 2). The methodology applied to future emissions is similar to that used for historical emissions, although with a lower sectoral and temporal resolution, as detailed below. This comprehensive gridded dataset is therefore produced with a consistent methodology by sector across all emission species spanning 1750 through 2014, with a consistent set of future projections over 2015–2100. The details of the methodology are first described for the historical emissions data (1750–2014), followed by a discussion of areas in which the methodology differs for future emissions (2015–2100). Further details including code are available online for both historical and future gridding (Sect. 5 below).

Aggregate anthropogenic emissions over the historical period by country and CEDS sector from Hoesly et al. (2018) are aggregated to 16 intermediate sectors (Table 2) and mapped to a $\mathrm{0.5}\times \mathrm{0.5}$${}^{\circ }$ grid as described below. The intermediate sectors were selected according to the availability of gridded proxy data in order to provide as much spatial detail as possible. For the final distribution, gridded emissions are further aggregated to nine sectors: agriculture, energy, industrial, transportation, residential–commercial–other, solvents, waste, international shipping, and aircraft. These final sectors were chosen to be consistent with previous practice (e.g., Lamarque et al., 2010), which includes matching the generally lower level of sectoral disaggregation available from future projections, and to reduce overall file size. Gridded emissions data are produced as CF-compliant NetCDF files (http://cfconventions.org/, last access: 4 February 2020) and converted to flux (kg m${}^{-\mathrm{2}}$ s${}^{-\mathrm{1}}$). Data were aggregated into a number of multiyear files, keeping individual file sizes under about 1 Gb. Filenames and NetCDF metadata are generated so that files are ready for distribution using input4MIPS on the Earth System Grid Federation system.

Emissions at the country and gridding sector level are mapped to a spatial grid using a variety of spatial proxy data, as described below. Emissions are distributed into target grid cells $XY$ as 1$${\mathrm{emission}}_{XY}={\mathrm{emission}}_{\mathrm{c}}\times {\displaystyle \frac{\mathrm{proxy}\mathrm{\_}{\mathrm{value}}_{XY}}{\sum \mathrm{proxy}\mathrm{\_}{\mathrm{value}}_{XY}}},$$ where ${\mathrm{emission}}_{XY}$ is the emissions value assigned to coordinate $XY$; ${\mathrm{emission}}_{\mathrm{c}}$ is the total country-level emissions for the gridding sector; $\mathrm{proxy}\mathrm{\_}{\mathrm{value}}_{XY}$ is the value of the proxy data for the cell $XY$, and the sum is performed over all coordinates $XY$ that are within the specified country.

The proxy data values are preprocessed for each country by multiplying the data by the area fraction of each grid cell that is in the specified country at an annual resolution (if annual gridded data are available). In this way grid cells that contain multiple countries will be assigned emissions proportionately. After assigning emissions to spatial grids by country, the spatially distributed emissions from each country are then added into a global spatial matrix. The “country” list includes a global region for emissions related to international shipping that are not associated with a particular country. Aircraft emissions are gridded separately, with one three-dimensional distribution that is scaled uniformly to match the global emissions estimate.

In most cases the proxy data are gridded emissions so as not to duplicate the effort needed to convert raw proxy data into the form needed for emissions inventories. We define two levels of spatial proxy data, the primary gridding proxies and a backup gridding proxy, which is gridded population. The backup gridding proxy is used where the primary gridding proxies are not appropriate due to either (1) the primary proxy not being available (e.g., is equal to zero) for the given country–sector–year combination or (2) the primary proxy being inaccurate for the given country–sector combination. In the latter case, we perform this proxy substitution when the ratio of proxy to sector emissions data for that country–sector combination is an outlier compared to the global distribution of this ratio across countries.

Over recent decades the primary gridding proxy data were from the EDGAR v4.2 (EC-JRC/PBL, 2012) inventory (Table 2) since these data were available over 1970 through 2008. Road transportation uses the EDGAR 4.3.2 road transportation grid, which is significantly improved over previous versions (Crippa et al., 2016) but was only available for 2010 at the time these data were produced, so the country-specific 2010 spatial distribution is used for all years. Flaring emissions use a blend of grids from EDGAR and ECLIPSE (Klimont et al., 2017). The backup gridded population proxy, as well as the proxy for early years for the residential–commercial sector, is a combination of gridded population from the Gridded Population of the World (GPW) (Doxsey-Whitfield et al., 2015) and HYDE (Goldewijk et al., 2011). Aircraft emission distributions are from Lee et al. (2009) and international shipping uses ECLIPSE shipping grids, with additional data from Endresen et al. (2003) for NMVOC emissions from oil tanker venting as used in Lamarque et al. (2010).

The only proxy data that vary before 1970 are those for the RCO (residential, commercial, other) and waste sectors (Table 2). For the RCO sector, for 1900 to 1969, the proxy linearly blends grids from the EDGAR v4.2 RCO 1970 grid and the gridded population, with the gridded population used for years before 1900. The proxy for the waste burning sector is based on rural population. More specific methodological details for certain aspects of the emissions dataset are outlined below.

The EDGAR and GPW proxy datasets that are used for most of the proxy data are initially processed at the highest resolution available (e.g., 0.1${}^{\circ }$ for the EDGAR emissions data) and split into countries, also including a split into land and ocean areas, at this resolution before being aggregated to the final 0.5${}^{\circ }$ resolution used for the data products. Because we allow multiple countries to exist within one 0.5${}^{\circ }$ grid cell, this results in a more accurate distribution of emissions for each country, even when data are being processed at the lower 0.5${}^{\circ }$ resolution. For the few datasets with a resolution lower than 0.5${}^{\circ }$, data are subsampled to a resolution of 0.5${}^{\circ }$ using an appropriate template.

The development of historical open burning emissions is described in van Marle et al. (2017). In brief, the spatial emissions distribution for these emissions is fundamentally grid-based, with emissions over the satellite era estimated from remotely sensed data (GFED4s; 1997–2015) merged with several existing historical proxies for fires, including charcoal records, for boreal and temperate North America and Europe as well as visibility-based fire emissions for the tropical areas of equatorial Asia and the arc of deforestation. The spatial distribution for the pre-satellite era was based on the 1997–2015 average and uniformly adjusted for large geographic regions based on the proxies (see van Marle et al., 2017, for the basis regions used). For the regions where proxies had limited coverage the output of six different fire models from the Fire Model Intercomparison Project (FireMIP) was used (Rabin et al., 2017) (https://www.imk-ifu.kit.edu/firemip.php, last access: 4 February 2020). Data are reported for 1750–2015. Before 1997 the annual regional fire emissions were distributed over a $\mathrm{0.25}\times \mathrm{0.25}$${}^{\circ }$ grid based on the GFED4s climatology for six classes (savanna and grassland fires, deforestation fires, boreal forest fires, temperate forest fires, peat fires, and agricultural waste burning).

Future open burning emissions originate from the integrated assessment models (IAMs), which report by model region and broad category (e.g., forest or grassland burning). While many of the models have some additional level of spatial detail, emissions were reported at the model region level to facilitate common data harmonization, downscaling, and gridding routines. Future open burning emissions were therefore mapped to spatial grids using the same methodologies as used for anthropogenic emissions, as further described below. This means that, unlike historical open burning emissions, the spatial distribution of open burning emissions within a category (e.g., forest burning) within any country does not vary in the future scenarios.

Scenario data for future emissions from IAMs are first harmonized to a common 2015 base-year value by native model region and sector. This harmonization process adjusts the native model data to match the 2015 starting year values with a smooth transition forward in time, generally converging to native model results (Gidden et al., 2018a). The production of the harmonized future emissions data is described in Gidden et al. (2019).

The 2015 base-year anthropogenic emissions data by country and sector are extensions of the 2014 historical data largely using emission factor trends for combustion sources from the GAINS model (ECLIPSE V5a; Stohl et al., 2015; Klimont et al., 2017) and BP fuel consumption statistics (BP, 2016). Noncombustion sources were generally scaled by estimated population. There are potentially large changes in emissions over this period, for example in China (Zheng et al., 2018), which results in uncertainty in these estimates for regions with rapid changes in air pollution control technology deployment. Near-term ${\mathrm{SO}}_{\mathrm{2}}$ emission trajectories in China were adjusted to better match the estimates of Zheng et al. (2018) during the harmonization process.

For open burning, the 2015 base-year emissions data for harmonization and downscaling are an average of the previous 10 years of historical data extracted by country. This is because interannual variability is not captured in the future projections, so a longer-term average is a more appropriate starting point for open burning emissions. While a decadal averaging period was used here, a longer averaging period could also be considered in future work.

The global integrated models (IAMs) that generated the Shared Socioeconomic Pathway (SSP) emission projections each have different numbers of socioeconomic regions (11–32) for which the unharmonized emission data are available. Harmonization to the common 2015 starting dataset therefore occurs at each individual model's native spatial definition in order to preserve as much detail from the model as possible (Gidden et al., 2019). Harmonized emissions for the native model regions are then downscaled to the country level as described in Gidden et al. (2019). The country-level downscaling is performed in order to provide a uniform basis for subsequent mapping to spatial grids but does not necessarily represent specific policies that might be in place for any particular country.

After the downscaling procedure the country-level future emissions projections are at the same level of sectoral resolution as the final gridding sectors in Table 2. These are then mapped to a spatial grid using the same underlying methodology as for the historical anthropogenic data, with a few differences in detail. Spatial proxies for each country are taken to be the 2014 gridded historical emissions discussed above for anthropogenic emissions and the average of the last 10 years (2005–2014) of historical data from van Marle et al. (2017) for open burning emissions. These spatial proxies are therefore constant into the future and do not represent any shifts in the spatial location of emissions within a country. The one exception is international shipping, which uses year-specific spatial distributions from the ECLIPSE project for the years 2015, 2020, 2030, 2040, and 2050. These distributions capture projected changes in shipping fuel sulfur content and the imposition of low sulfur control near coastal areas.

Future open burning emissions are provided from the IAMs in the following categories: agricultural waste burning on fields (AWB), forest burning, grassland burning, and peatland burning. Future trends for these emissions are generated by each model driven by, in large part, changes in land use. Note that these are future projections of climatologically average emission rates over time and do not include interannual variability. Because the future model projections used here do not include data on peatland burning, peatland burning emissions are held constant into the future.

Future emissions were gridded at the aggregate sectoral level, which corresponds to the final gridding sectors in the historical emissions data for anthropogenic emissions (Table 2), and with forest burning combined into one category, aggregating the open burning historical categories of boreal forest fires, temperate forest fires, and tropical deforestation and degradation. This lower level of sectoral detail was used because the models that generated the future scenario data used in this process (Gidden et al., 2019) often lack the finer level of sector detail that was available in the historical emissions datasets. The future data were constructed so that the data were consistent at the grid cell level when moving from the historical (up to 2014) to the future (2015 and forward) dataset.

The second major difference is that future emissions are not produced annually but are provided for 2015, 2020, and at decadal intervals thereafter. This is because long-term models do not provide annual data. The choice was made to only distribute data for years that were originally produced by the models and not interpolated data. This also means that fewer, and smaller, data files need to be downloaded and processed by end users. We note that the format of the open burning emissions in the future data is slightly different than the format used in the historical data. This is because we use the same software for both open burning and anthropogenic future emissions, so a similar data format was used for all future emissions data.

Monthly seasonality is applied to the gridded emissions by applying a set of spatially explicit seasonality fractions. The primary source for the seasonality fractions was the ECLIPSE dataset, with the addition of EDGAR v4.3 for international shipping and Lamarque et al. (2010) for aircraft. The dataset was processed to a consistent seasonality that matched the monthly calendar used in the final emissions data (e.g., a 365 d year). As part of this process some minor inconsistencies in the anthropogenic seasonality data between sectors in terms of the number of days assumed in the year were corrected so that the month distribution was consistent with the 365 d year used in the anthropogenic dataset. Seasonality for future open burning emissions was taken from a 10-year average from the historical dataset.

We note that, after the distribution of this dataset, it was found that the industrial sector in some regions has a high level of seasonality in the ECLIPSE dataset, and this distribution was carried through to the CMIP6 data (as noted on the online GitHub issues list for this dataset; see below). In general, we would expect industrial sector emissions to be fairly constant as most industrial activities operate year-round. The magnitude of seasonality overall in some regions may therefore be overestimated. This will be reexamined in future data releases.

Modeling atmospheric chemistry requires emissions for specific species, or species groups, of reactive compounds instead of the total mass of volatile hydrocarbon emissions (NMVOCs) generally tabulated in inventories. Anthropogenic VOC emissions were speciated by applying speciation profiles by gridding sector and country, largely derived from the RETRO project. Emissions for 23 anthropogenic VOC groups were extracted by country and broad sector from gridded HTAP v2 emissions data (http://iek8wikis.iek.fz-juelich.de/HTAPWiki/WP1.1, last access: 13 January 2016). These were normalized, converted to percentages, and applied to the emissions data at the country and final gridding sector level. In the future, it would be useful to use country- and sector-specific profiles, such as those in Huang et al. (2017). The speciated emissions were gridded in the same manner as bulk VOC emissions, thereby providing emission grids by species.

While speciation profiles by gridding sector are held constant in time, the aggregate VOC speciation for a country will change over time. This is because of different speciation profiles for each sector combined with a changing sectoral contribution to total VOC emissions over time.

Figure 2 shows the ratio of butanes to hexanes$+$ in the RETRO dataset used in CEDS and the newer estimate by Huang et al. (2017), which we present as an illustration of differences in the speciation profiles used in these two datasets. While there is considerable overlap between the two datasets, the RETRO data have a lower average ratio for these two species in the road transportation sector compared to EDGAR, while the opposite is the case for industrial combustion. We note that variation across, and even within, countries is expected. For example, Chin and Batterman (2012) report that the fraction of $n$-heptane (a higher alkane) in gasoline vapors varies between 6 % and 19 %, depending on location and season within the United States.

For both historical and future emissions, however, VOC speciation profiles can potentially change with technology deployment separately from changes in NMVOC emission rates. For example, regulations to limit ozone formation not only result in altered bulk NMVOC emission amounts, but also change the specific VOC species emitted (Kirchstetter et al., 1999). Other examples include changes over time in the composition of consumer products (McDonald et al., 2018) and changing formulation of paints and other coating materials.

It is not known how much the differences illustrated in Fig. 2 reflect updated or different data sources or actual changes in VOC speciation over time. The RETRO dataset is targeted at 1990–1995 and uses speciation profiles available in time period, while the EDGAR data take in newer measurements. While the emission time series presented here, and also the EDGAR data, capture the broad changes in VOC speciation due to changing sectoral contributions to VOC emissions over time, these datasets do not capture underlying changes in speciation profiles due to changing regulations or other fundamental changes to speciation profiles over time. It would therefore be useful to determine the importance of such speciation changes for the historical modeling of atmospheric chemistry compared to the broader changes in VOC emission magnitudes and speciation changes due to sectoral shifts over time. Note that the Huang et al. (2017) data do capture changes in speciation over time due to changing mixes of fuels in each end-use sector, which is illustrated in Fig. 2.

For future open burning emissions, speciation profiles by country and open burning sector were extracted from the historical open burning emissions data files and applied to the future scenario bulk NMVOC emissions. For details on the historical open burning VOC speciation see van Marle et al. (2017). Note that the species provided are different for the anthropogenic emissions and open burning emissions, following the past conventions used by each of those communities.

Table 4 provides a summary of the gridded data files described in this paper with the emission species listed in Table 1. The files listed there, plus historical gridded open burning emissions (van Marle et al., 2017), provide the complete set of emissions data required for CMIP6 experiments (Eyring et al., 2016). As discussed in Hoesly et al. (2018) the anthropogenic emissions are consistently generated from the same driver data across sectors and emission species. The gridded data files are generated using consistent spatial proxy data, seasonality, and sector definitions. This means that some emission species, such as ${\mathrm{CO}}_{\mathrm{2}}$, which have been provided as only bulk emissions in CMIP5, are provided here with greater sectoral detail. For applications that do not require this level of sectoral detail, emissions can be summed across sectors. Note that, for all species, total anthropogenic emissions consist of the sum of all eight (historical) or nine (future)

As discussed above, future emissions contain an additional sector to represent net negative ${\mathrm{CO}}_{\mathrm{2}}$ emissions.

Supplementary data files include speciated NMVOC emissions for all time periods. Also provided are historical emissions from solid biomass fuels (wood, agricultural residues, etc.) as these are used as supplementary data in some models.

Some models use information on solid biofuel combustion to derive information on associated VOC emissions and/or primary aerosol (e.g., BC, OC) characteristics.

Spatial distributions for anthropogenic reactive gases and carbonaceous aerosol emissions (e.g., BC and OC) for selected historical and future years are shown in Figs. 3–6, with numerical values in Table 3. An overview of large-scale trends is provided here, with the emission trajectories and their derivation described in more detail in the underlying literature for the historical emissions (van Marle et al., 2017; Hoesly et al., 2018), future scenarios (Calvin et al., 2017; Fricko et al., 2017; Fujimori et al., 2017; Kriegler et al., 2017; Riahi et al., 2017; van Vuuren et al., 2017), and the harmonized CMIP6 data (Gidden et al., 2019).

Global emissions of species that are the product of incomplete combustion (BC, CO, ${\mathrm{NO}}_x$, OC, NMVOC) are dominated by open burning in 1850 (Table 3, Figs. 3–4). Anthropogenic combustion emissions in 1850 are dominated by solid biofuel use for cooking and heating. However, by the mid-20th century, anthropogenic processes dominate for all species except for OC (Table 3). As industrialization proceeds, emissions become large in the regions that experienced early industrialization such as Europe, North America, Japan, and Australia by the middle to late 20th century (e.g., Fig. 5). By the early 21st century (e.g., 2015; Fig. 6) emissions controls have decreased pollutant emissions somewhat in Europe and North America but increased dramatically compared to mid-20th century levels in emerging economy regions such as China, India, and the Middle East. Anthropogenic ${\mathrm{CO}}_{\mathrm{2}}$ emissions, in general, do not follow the same trajectory as air pollutant emissions.

Future emissions diverge depending on the emission scenario and follow trends as described in Gidden et al. (2019). A few illustrative graphics are provided here. By 2050 in SSP2, the “middle of the road” scenario, emissions have decreased from 2015 levels in most world regions (Fig. 7), while they have decreased to quite low levels in the SSP1-26 scenario (Fig. S7 ). In contrast, emissions do not decrease nearly as much in the SSP3-70 scenario (Fig. S9). By 2100, under the SSP2-45 scenario, anthropogenic emissions are generally larger than open burning emissions, but anthropogenic $+$ open burning emissions are generally smaller than 2015 levels. Future open burning emissions also vary by scenario, but not as dramatically as anthropogenic emissions (Figs. S3–S6 in the Supplement).

While 0.5${}^{\circ }$ gridded data are the primary product of gridding, maps at that resolution can be of limited use for understanding spatial and temporal patterns in the data or diagnosing potential issues. In this section we discuss a number of diagnostic graphs that illustrate various dimensions of the dataset. We first discuss the temporal structure of the anthropogenic emissions data over different sectoral and spatial scales.

Figure 8a shows a time series plot of ${\mathrm{SO}}_{\mathrm{2}}$ emissions at one grid cell for the industrial (IND) gridding sector (e.g., combustion $+$ process emissions; see Table 2). These emissions are the result of a combination of changes in the proxy data over time, in this case from EDGAR emissions, multiplied by the estimated time series for the United States for the underlying detailed sectors. Emissions from this same cell for the energy sector (Fig. 8b) show a sharp increase due to (presumably) a power plant coming into service, as represented in the proxy dataset.

Total anthropogenic ${\mathrm{SO}}_{\mathrm{2}}$ emissions for the same grid cell (Fig. 8c) show a different temporal structure as a result of combing the differing temporal trends from different sectors. Emissions become smoother if we examine total ${\mathrm{SO}}_{\mathrm{2}}$ emissions summed over 16 nearby grid cells (Fig. 8d). However, emissions at either the grid cell or for a sum of several grid cells will generally show more structure over time compared to the corresponding emissions for the entire USA region (Fig. 9) due to the additional temporal structure in the proxy dataset at smaller spatial resolutions.

For any particular spatial location, the accuracy of the gridded emissions depends on the accuracy of the underlying proxy dataset as well as underlying the country-level emissions data. Global proxy datasets that are available for use in projects such as this, including projects of similar scope such as EDGAR (from which many of the spatial distributions used here are drawn), will not necessarily correspond exactly to the actual spatial location and magnitude of emission sources. For example, databases of global road networks, while they are improving (e.g., Crippa et al., 2018), will not exactly correspond to either actual road traffic at every point, nor are they likely to be complete for all regions. While large emission sources, such as power plants, are more likely to be in global databases, such data can still be incomplete, particularly for smaller plants (Liu et al., 2015). For other sectors, it is difficult to capture in proxy datasets historical changes over time in emission strengths in specific spatial locations.

At the grid cell level, discontinuities over time can therefore potentially be due to (a) discontinuities in the country-level emission time series, (b) actual changes in the underlying emitting processes as represented in the proxy dataset, or (c) data breaks or inconsistencies in the proxy dataset that do not represent an actual historical change in emissions. As either the sectoral aggregation or the spatial scale increases, however, the data will tend to become more robust, at least to the extent that the country-level emission time series used for calibration in this project are accurate (see Hoesly et al., 2018, for a further discussion).

Monthly time series plots visualizing the seasonal cycle of the emissions data, as shown in Fig. 10, are also useful diagnostics. Note, again, that discontinuities at the single cell level are due to temporal structure in the proxy datasets.

The temporal signature of the open burning data is quite different, as interannual variations in those data are dominated in large part by year-to-year changes in local meteorological conditions. At least for the modern era, the interannual variations are inferred from satellite observations. As discussed below, the increased use of remote sensing and regional bottom-up data has the potential to improve the spatial and temporal accuracy of anthropogenic emissions data as well.

Spatial data presented at a lower resolution can also be useful in understanding overall patterns and spatial differences between datasets. As discussed in Hoesly et al. (2018), we have found that spatial maps at a 10${}^{\circ }$ resolution are useful in showing large-scale differences in emissions magnitude and spatial distribution between two datasets. For anthropogenic ${\mathrm{SO}}_{\mathrm{2}}$ emissions in 1900, for example (e.g., Fig. 11), the CMIP6 dataset shows some shifts in emissions from northwest to eastern Europe and a different distribution of emissions over the oceans compared to the CMIP5 dataset.

Also note that the CMIP6 dataset has a different distribution of ${\mathrm{SO}}_{\mathrm{2}}$ emissions over North America than in the CMIP5 data, with more ${\mathrm{SO}}_{\mathrm{2}}$ emissions over the western areas and lower emissions over the east. This has been recognized (Yang et al., 2018) as a bias in the proxy dataset over the United States used here (e.g., EDGAR ${\mathrm{SO}}_{\mathrm{2}}$ emissions). As discussed above, national-level anthropogenic emissions were distributed within each country using proxy data. In this case, the proxy data likely do not account for the lower sulfur content of coal in the western United States. As discussed below, one method of accounting for such issues is to estimate emission at smaller geographic units, such as US states, and then map to spatial grids. For further comparisons between the CMIP5 and CMIP6 anthropogenic emissions data, see Hoesly et al. (2018).