Air pollution is a crucial environmental concern because of its detrimental impacts on health, ecosystems, the built environment, and short-term climate forcing. Whereas it was originally regarded as an urban issue, in the late 1970s the large-scale acidification of precipitation made it clear that at least part of the problem could only be solved through international cooperation (OECD, 1977). This was the background for the establishment of the Convention on Long Range Transboundary Air Pollution (CLRTAP) in 1979. The main vehicles of the LRTAP Convention are the protocols that aim to reduce the emission of various compounds (sulfur in 1985, nitrogen oxides in 1988, volatile organic compounds in 1991, heavy metals and persistent organic pollutants in 1998, and the multi-pollutant multi-effect Gothenburg Protocol to abate acidification, eutrophication, and ground-level ozone in 1999, with subsequent revision in 2012). The design of such mitigation strategies was largely supported by the development of models (chemistry-transport and integrated assessment tools) and monitoring networks.

After several decades of international cooperation, it is timely to take stock of the evidence available to assess the efficiency of the LRTAP Convention and the corresponding emission-ceiling protocols. The executive body of the Convention has therefore requested an assessment of the evolution of air pollution and subsequent effects from its two scientific and technical bodies: (i) the European Monitoring and Evaluation Programme (EMEP) and (ii) the Working Group on Effects (WGE). As a result, the Task Force on Measurement and Modelling (TFMM) of EMEP published an assessment of air pollution trends (Colette et al., 2016), whereas the WGE published an assessment of corresponding effects on health and ecosystems (De Wit et al., 2015), and an overall assessment report encompassing all the activities undertaken under the Convention was also released (Maas and Grennfelt, 2016).

The effects of emissions on the concentrations is rather complex due to (i) the non-linearity of atmospheric chemistry, (ii) the presence of inflow of air pollution due to the intercontinental transport of air pollutants, and (iii) the meteorological variability. This is where chemistry-transport models (CTMs) come into play with the multi-model air quality trend experiment introduced in the present paper.

The LRTAP convention relies in part on the results of the EMEP/MSC-W chemistry-transport model (Simpson et al., 2012a). Since the beginning of the 2000s, the Joint Research Centre of the European Commission initiated a number of multi-model assessments to provide a benchmark for the EMEP/MSC-W model through its comparison with the modelling tools being used by the states and parties to the Convention as part of the Eurodelta project (Bessagnet et al., 2016; van Loon et al., 2007; Thunis et al., 2008). The EURODELTA-Trends (EDT) exercise builds upon this tradition, focusing on the specific context of air quality trend modelling. Its main goal is to assess to what extent observed air pollution trends could be related to emission mitigation, although this overarching question can only be addressed after having assessed the confidence we can have in the models, and in particular in their capacity to reproduce the trends.

In the recent past, several multi-model projects covering a time period of 1 year or less were undertaken, such as the earlier phases of Eurodelta cited above but also the various phases of the AQMEII project (Galmarini et al., 2012, 2017; Rao et al., 2011; Im et al., 2015). However, only a few attempts have been made to address the issue of the long-term evolution of European-scale air quality by means of modelling studies. The first attempts were using only one model as in Vautard et al. (2006), Jonson et al. (2006), and Wilson et al. (2012). A first ensemble was proposed through the European Project CityZen which relied on six models (Colette et al., 2011). While these studies were limited to about 10-year time periods, a 20-year hindcast study was presented in Banzhaf et al. (2015), relying again, however, on a single model. It is therefore timely to engage in a multi-model hindcast of air quality over 2 decades.

The purpose of the present paper is to define the science and policy questions that are addressed by the EDT exercise, and introduce the experimental setup designed to answer these questions. The models participating in the experiment will also be presented, as well as the project database of model results.

The main policy focus being addressed in EDT analysis is the assessment of the role of European air pollutant emission reductions in improving air quality over the past 2 decades. Subsequent questions include assessing (1) the role of changes in global air pollution as well as (2) the role of inter-annual meteorological variability. Before addressing such issues, it will be essential to quantify the CTMs' capability in (1) reproducing observed air pollutant concentrations (processes determining air quality: chemistry, physics, transport processes, emissions, meteorology) and (2) capturing the long-term evolution of air quality.

The time period covered by the experiment is 1990–2010. The year 1990 has been chosen as the beginning of the period because that year serves as reference for the Gothenburg Protocol. The end of the period is 2010 because of the availability of underlying forcing data (emissions, boundary conditions (BCs), and meteorology) required for model calculations at the time the work was initiated.

The EDT model experiment is divided into three tiers, targeting various science and policy questions. The tiers also differ in terms of computational demand that allowed the involvement of as many modelling groups as possible. The tiers of experiments are summarised in Table 2. They differ in terms of the number of modelled years to be addressed in the 1990–2010 period and in terms of forcing data used in model calculations for the anthropogenic emissions, the chemical boundary conditions, and the meteorological year. Most of the experiments consist of variations in one or two of these three factors in order to disentangle the role of each forcing. The role of chemical boundary conditions constitutes one notable exception, since two sources of forcing are used: either a global CTM simulation or an observation-based climatology (further details are provided on boundary conditions in Sect. 7).

The first simulation of the EDT experiment is a reference for the year 2010 using the meteorology (M), the chemical boundary conditions (B) and the emissions (E) for that year, named as M10B10E10, with two digits corresponding to the last two digits of the year. They are complemented with simulations for the years 1990 and 2000 (using corresponding meteorology, boundary conditions, and emissions: M90B90E90 and M00B00E00, respectively) to form tier 1A. Tier 1A will allow testing of the accuracy of all CTMs in simulating pollution changes for the near past (1990, 2000, and 2010), at a lower computational cost than running the full 21-year period.

Tier 1B is dedicated to the first two sensitivity experiments, for which the meteorology and the boundary conditions are those of the year 2010, but the emissions correspond to 1990 and 2000 (M10B10E90 and M10B10E00). They will allow assessment of the individual impact of emission changes alone (E10 vs. E90 and E10 vs. E00) by comparison with tier 1A (specifically M10B10E10).

In tier 2A, two more sensitivity simulations are performed for the meteorological year 2010, using emissions and boundary conditions of 1990 and 2000 (M10B90E90 and M10B00E00, respectively). By comparison with tier 1B, they will allow the assessment of the impact of global chemical background changes on European air quality between the years 1990 and 2010, and also for the sub-periods 1990–2000 and 2000–2010 (B10 vs. B90 and B10 vs. B00).

Tier 2B is an alternate set of reference simulations for 1990, 2000, and 2010, in which boundary conditions provided by a global model (C) instead of the observation-based boundaries are used (M90C90E90, M00C00E00, M10C10E10). It will allow assessment of the uncertainty related to the large-scale chemical forcing by comparison with tier 1A.

Tier 2C is a complement to tier 2A using the meteorology of 2000 and two combinations of 1990 and 2000 boundary conditions and emissions (M00B90E90, M00B00E90). These additional simulations are required to perform the attribution analysis for the concentration changes between 1990 and 2000, whereas the simulations required for the attribution of driving factors between 1990 and 2010 and between 2000 and 2010 are dealt with in tiers 1A, 1B, and 2A.

Tier 3A consists of 21-year simulations covering 1990–2010, using meteorology, boundary conditions and emissions for the respective years (MyyByyEyy, with yy being the 2-digit year between 1990 and 2010). It will be used to assess the capability of the models to capture observed trends in air quality by means of comparisons with available measurements. Fewer modelling teams delivered results for this higher tier of experiments, and therefore model uncertainty will be put in perspective with the spread of the whole ensemble in modelling tier 1A (1990, 2000, and 2010).

Tier 3B is the last sensitivity experiment in which 21-year simulations are performed using the 2010 emissions for the complete period (MyyByyE10, with yy varying from 90 to 10, refering to the years 1990 to 2010). By comparison with tier 3A, it will allow the determination of the role of inter-annual meteorology and chemical boundary condition changes vs. the role of European emission changes.

Thus, the complete series of model runs included for each air quality model is 5 annual simulations for tier 1, 7 more simulations for tier 2, and 39 (2×21 minus 1 overlap for 2010, and 2 annual simulations belonging to tier 1A: M90B90E90 and M00B00E00) more simulated years for tier 3.

Figure 1 provides the schematics of the various combinations of simulations required to perform the attribution analysis for any period of time between the three reference years (1990, 2000, and 2010). The simulations labelled in black are covered by the above simulation plan. They are needed for the assessment of the relative role of emissions, meteorology, and boundary condition changes.

The main limitations of the simulation plan are that (i) the three selected meteorological years may be not representative, or atypical, for the full period and (ii) the lack of interaction by considering 22 combinations instead of the 23 combinations required to cover the whole space of factors (Stein and Alpert, 1993). In the forthcoming attribution study these limitations will be explored by (i) comparing trend (tier 3A) and sensitivity (tier 1 & 2) tiers and (ii) including additional simulations for the 23 possible combinations from one of the models (CHIMERE).

Eight European modelling teams submitted their calculation results to the EDT database for at least one tier of experiment (see the experiment design in Sect. 2) using state-of-the-art air quality models: CHIMERE (Menut et al., 2013; Mailler et al., 2017), CMAQ (Byun and Schere, 2006), EMEP/MSC-W (Simpson et al., 2012), LOTOS-EUROS (Schaap et al., 2008; Manders et al., 2017), MATCH (Robertson et al., 1999), MINNI (Mircea et al., 2016), Polyphemus (Mallet et al., 2007), and WRF-Chem (Grell et al., 2005; Mar et al., 2016). The main specifications of the eight participating models are summarised in Table 1 (note that they can differ from the public release of the various models according to the elements provided in the table).

The representation of physical and chemical processes differs in the models. The vertical distribution of model layers (including altitude of the top layer and derivation of surface concentrations at 3 m height in the case of EMEP, LOTOS-EUROS, and MATCH) is not prescribed either. However, as further explained in the article, the other features of the model setup are largely constrained by the experiment input data such as forcing meteorology, boundary conditions, emissions, and the experiment characteristics such as horizontal domain and resolution. Only one of the participating models included online coupled chemistry–meteorology (WRF-Chem), while all the other models are offline CTMs.

The modelling domain is displayed in Fig. 2. The domain follows a regular latitude–longitude projection (plate carrée projection) with increments of 0.25 and 0.4∘ in latitude and longitude, respectively, which is about 25 km × 25 km at European latitudes. The total coverage extends from 17∘ W to 39.8∘ E and from 32 to 70∘ N. All the participating models use the same modelling domain, with only one exception: CMAQB uses a Lambert conformal conic projection map with 25 km resolution and delivered their results on the common grid. The southeasternmost part of the domain is not included in the CMAQB modelling domain.

The horizontal resolution of available global meteorological reanalyses over the 1990–2010 period is considered too coarse to drive regional-scale CTMs. Therefore, dynamically downscaled regional climate model simulations using boundary conditions from the ERA-Interim global reanalyses (Dee et al., 2011) were used to force the CTMs involved in EDT. Most CTMs used the same meteorological driver, with a couple of exceptions.

One of the meteorological drivers was produced using the Weather Research and Forecast Model (WRF version 3.3.1; Skamarock et al., 2008) at 0.44∘ of resolution. In the framework of the EuroCordex climate downscaling programme (Jacob et al., 2013), an evaluation of the regional climate models downscaled with reanalysed boundary conditions (ERA-Interim reanalyses instead of free climate runs) was reported by Kotlarski et al. (2014). One of the WRF realisations in the EuroCordex ensemble was subsequently further optimised as described in Stegehuis et al. (2015), so that we could identify an optimal WRF setup for our purpose (row no. 7 of Table S1 in their Supplement). The model was re-run using grid-nudging towards the ERA-Interim reanalyses (above the planetary boundary layer) in order to improve temporal correlations compared to the regular free-running Cordex hindcast simulations. This WRF simulation was interpolated on the 25 km resolution EDT grid and used to drive CHIMERE, EMEP, Polyphemus, and MINNI. In the EMEP model, the interpolation of the meteorological fields from 0.4×0.4∘ to EDT grid was performed online. For WRF-Chem, an online model that simulates meteorology and chemistry simultaneously (“online”), the meteorology from the WRF-Eurocordex runs (Stegehuis et al., 2015) was used as initial and lateral boundary conditions and for applying four-dimensional data assimilation (FDDA), with coefficients as described in Mar et al. (2016). The CMAQ model, which runs on a Lambert conformal conic projection, could not use the meteorological data provided on the EuroCordex grid, so that WRF was re-run in a Lambert conformal projection at 25 km horizontal resolution using identical WRF setup and version (3.3.1).

The CTMs LOTOS-EUROS and MATCH have been meteorologically forced by ERA-Interim series further downscaled with RACMO2 (van Meijgaard, 2012) and HIRLAM (Dahlgren et al., 2016), respectively. RACMO2, used here, was part of the EuroCordex studies documented in Jacob et al. (2013) and Kotlarski et al. (2014) and excludes nudging towards Era-Interim. The HIRLAM EURO4M reanalysis uses data assimilation in three dimensions for upper air and optimal interpolation for surface fields. An initial analysis is conducted every 6 h with subsequent forecasts saved on 3-hourly temporal resolution. ERA-Interim is forced to the lateral boundaries. The HIRLAM reanalysis was interpolated from the original 0.2 horizontal resolution on a rotated latitude–longitude grid (ca. 22 km) to the EDT grid. The main features of the meso-scale meteorological models are synthesised in Table 4.

Two sources of lateral and top chemical boundary conditions are used by the regional CTMs: a climatology of observational data and global model results. Both have pros and cons. Global models carry biases but include a wider array of chemical species. The trend in observations matches in situ data by nature, but only at one point over the domain. For the EDT experiment the consensus in the experiment design was in favour of observation-based boundary conditions for most experiments (tier 1A, 1B, 2A, 2C, 3A, 3B) but also includes a sensitivity study based on modelled boundary conditions (tier 2B).

Note that a possible impact of changing chemistry composition on large-scale circulation was integrated in the forcing meteorological fields through the data assimilation of the ERA-Interim reanalysis. This factor was not considered important to isolate for the 2-decade timescale of the experiment.

Note also that both sources are provided on the basis of monthly averages so that sporadic advection of large intercontinental pollution plumes or dust events will not be captured, although their impact on monthly means is taken into account.

The model simulations were delivered in a common NetCDF format, so that each of the files contains gridded fields of one pollutant for a whole year. The air pollutant concentrations from only the lowest model level (or corrected to 3 m height for EMEP, LOTOS-EUROS, and MATCH) are delivered to the project database, but the participants are encouraged to store three-dimensional data if their storage capacities allow such an archiving.

The requested variables are as follows:

Hourly concentrations of O3 (O3_HL) and NO2 (NO2_HL);

Additional diagnostics were subsequently computed and delivered on the common database; the list of indicators and their definitions is available in Table 5.

The status of models' delivery of results for each of the experiment tiers at the time of submission of the present article is summarised in Table 3. The access to the database is open for research use through the AeroCom server (see also the section on data availability)

https://wiki.met.no/aerocom/user-server (last access date 14 June 2017)

A few illustrations of the results of the Eurodelta-Trend multi-model air quality hindcast are provided in Figs. 4 and 5 with ensemble-median and ensemble-spread concentration maps of ozone and particulate matter in 1990 and 2010, obtained from the eight models which delivered output to tier 1A. For ozone, we show the summertime (June–July–August) average of the daily maxima of 8 h mean ozone. For particulate matter, the annual mean PM10 is presented.

It is the ambition of the whole Eurodelta-Trend experiment to assess how those maps compare with observations, both in terms of spatial variability and temporal trends, and also to further explain the rationale for the changes observed between 1990 and 2010. However, such analyses require substantial work that is left out of the present article devoted to the presentation of the experiment.

It is worth highlighting, however, that substantial decreases of both ozone peaks and particulate pollution are modelled in the EURODELTA-Trends ensemble between 1990 and 2010. We present here the decrease on the basis of 1990 and 2010 snapshots for the whole eight-model ensemble that contributed to the experiment. But it would require further documentation in terms of trends by comparison with the subset of five models that produced the full set of 21-year trend simulations in tier 3A.

For summertime ozone, concentrations exceeding the European target value of 120 µg m-3 are only found in the greater Mediterranean region in 2010, whereas in the early 1990s, such concentrations affected much larger areas of continental Europe. The spread (standard deviation) of the models is much larger in 1990 than 2010, especially over the polluted areas of Europe at that time.

Particulate matter concentrations also decreased substantially. The largest spread in the eight-member ensemble is found over sea and desert areas (also because absolute concentrations are high over northern Africa), where the differences between the models changes significantly between 1990 and 2010. This raises important questions regarding the uncertainties of the models for natural sources and the role of inter-annual meteorological variability on aerosol concentrations.

The Eurodelta-Trend modelling experiment will allow a better understanding of the evolution of regional-scale air quality over Europe over the 1990–2010 period. This is facilitated by the thoroughly designed modelling plan. Eight modelling teams have participated in the EDT experiment, though with a variable degree of involvement. The base runs of tier 1A, completed with eight participating models, offer a great opportunity to assess the capability of these state-of-the-art chemistry-transport models to reproduce the observed changes in the concentrations of the main pollutants, including ozone, particulate matter, and its individual components, as well as in precipitation chemistry. This analysis will then be complemented by an assessment of the capability of reproducing the actual trends over the 21 years in the 1990–2010 period for the models participating in the more demanding tier 3A experiment. If this evaluation phase concludes that the skill of these models in capturing air quality evolution is satisfactory, we would then rely on the results of the trend (or decadal-change) calculations and the sensitivity experiments and recommend that they can be used when addressing science and policy questions underlying the evolution of air quality in Europe over the past couple of decades.

The critical policy question lies in the attribution of air quality trends to emission changes, to influx at the boundaries of the European domain, and to interannual meteorological variability (and natural sources of trace species) and will be addressed in a series of upcoming papers. Furthermore, thanks to the multi-model design of the experiment, other scientific questions with regard to the role of specific chemical and physical processes will be investigated in forthcoming studies based on the EURODELTA-Trends results.

The model results will also be publicly distributed in order to serve for in-depth analyses to scientific communities working on the impacts of air pollution on health, ecosystems, or aerosol radiative forcing.