The anthropogenic influence on atmospheric chemistry is apparent at all spatial and temporal scales: human emissions have impacted local and very short-lived species (e.g., OH; see Prinn et al., 2001), very long-lived greenhouse gases (e.g., Collins et al., 2006), and everything in between (e.g., Baker et al., 2015; Solomon et al., 2016). Over the past decades, all three branches of modern atmospheric chemistry research (Abbatt et al., 2014) – observations, laboratory analysis, and modeling – have increased in both their sophistication and their capability to explain the chemistry of our atmosphere. However, while observational networks have significant growth potential (e.g., Sofen et al., 2016) and laboratory analysis still has significant challenges to overcome (Bocquet et al., 2015; Burkholder et al., 2017), chemistry modeling efforts are finding their growth potential limited by the level of chemical complexity that can be included in models due to the constraint of the computational capabilities of even state-of-the-art supercomputers (Stockwell et al., 2012). Simulations that attempt to include all known species and reactions, such as the National Center for Atmospheric Research (NCAR) Master Mechanism (Madronich and Calvert, 1989; Aumont et al., 2000) or the Leeds Master Chemical Mechanism (Jenkin et al., 1997; Saunders et al., 2003) and even some species and reactions that have not been tested in any laboratory (e.g., Aumont et al., 2005; Szopa et al., 2005), are often limited to box-model-level analysis (e.g., Emmerson and Evans, 2009; Squire et al., 2015). Modeling efforts that simulate regional- and global-scale atmospheric chemistry are forced, out of practical necessity, to utilize simplified, reduced-form, and parameterized chemistry in order to address the large spatial and long temporal scales needed for policy-relevant research.

Historically, as computational capacity has increased, modeling efforts have tended to maximize model resolution and complexity. This limits the capability to perform multi-scenario or multi-model ensembles to institutions with access to significant computational capabilities and storage. One way to increase the number of scenarios or members in an ensemble is to reduce the complexity of the chemical mechanism. This selection of a reduced-form chemical mechanism for different applications and the advantages of the increased computational efficiency of a simplified mechanism are the main focus of this paper. While there is a long history of publications (see Dodge, 2000) that compare different photochemical mechanisms within box models (e.g., Milford et al., 1992; Jimenez et al., 2003; Emmerson and Evans, 2009; Knote et al., 2015), studies that compare multiple mechanisms within a single 3-D global model are rare (e.g., Squire et al., 2015). This study examines three chemical mechanisms within the Community Earth System Model Community Atmosphere Model with Chemistry Version 1.2 (CESM1.2 CAM-chem; Lamarque et al., 2012) framework: the Model for Ozone and Related Chemical Tracers, version 4 (MOZART-4) mechanism, the Reduced Hydrocarbon mechanism, and the Super-Fast chemical mechanism (described in Sect. 2), which is one of the simplest representations of atmospheric chemistry used within climate–chemistry model intercomparison projects, such as the Atmospheric Chemistry and Climate Model Intercomparison Project (ACCMIP; Lamarque et al., 2013).

This study examines the trade-offs and possibilities that arise from the selection of a chemical mechanism that is simple enough to be computationally efficient – and thus capable of long simulations or large ensembles at the global scale – as well as sophisticated enough to simulate the major features of tropospheric chemistry at the local and regional scale. Many climate studies include little to no chemistry or prescribed chemistry, even though chemistry–climate feedbacks are well established to impact global and regional climate (e.g., Marsh et al., 2013; Fiore et al., 2015). Indeed, coarse-grid (2∘×2.5∘) chemistry–climate studies which conduct 1000 or more years of simulations using complex chemistry are notable in their rarity (see Barnes et al., 2016, and Garcia-Menendez et al., 2015, 2017). This paper focuses on three primary lines of inquiry focusing on tropospheric ozone. First, what is lost or gained with the selection of a simplified chemical mechanism within a global model? Second, what is the nature of the uncertainties that arise with the selection of a particular chemical mechanism? And third, what are the trade-offs that researchers make, either intentionally or tacitly, when they apply a specific mechanism within a particular modeling framework? We focus this study on the short-lived gaseous species, in particular ozone and its precursors, that influence both the daily exposure of humans to pollutants as well as the decadal-scale global climate system. We focus primarily on a computationally efficient simulation of tropospheric gaseous chemistry within a single modeling framework and leave further analysis of other aspects of atmospheric chemistry to future studies.

In Sect. 2, we describe the modeling framework and describe each of the three aforementioned chemical mechanisms, including a detailed description and history of the Super-Fast mechanism, as it is not reported elsewhere in the literature, and the simulations and observations we use for comparison. In Sect. 3 we present spatial and temporal results and compare various metrics of chemical accuracy. In Sect. 4, we explore the nature and the morphology of the chemical uncertainties and the particular trade-offs that are made by the selection of a single mechanism when faced with limited computational resources. We draw conclusions in Sect. 5.

Our analysis focuses on characterizing the ozone chemical uncertainties within a global chemistry model. We examine the morphology of the chemistry system, focusing specifically on the means, standard deviations, and variability (defined here as the standard deviation divided by the mean). We also include characterizations of the correlation of the ozone time series with the observations and of the extreme values (in particular the 90th and 99th percentiles) of the ozone distribution.

Our primary objective has been to determine what is lost (or gained) with the selection of a simplified chemical mechanism, which we summarize here. We mostly discuss the SF mechanism, as the trade-offs with the RH mechanisms are straightforward: we lose very little (Fig. 10a and i) and gain about a 100 % increase in simulation speed (Table 1). Many of the things that are lost with the use of the SF mechanism are expected: we lose the capability to directly simulate small-scale features of ozone chemistry in regions that depend strongly on complex biogenic chemistry. In particular, the equatorial landmasses – especially equatorial Africa and South America – are not well simulated (Fig. 7). We also lose the capability to simulate some of the short-term features that require additional chemistry, such as the nighttime behavior of isoprene (Fig. 9) or the cold season CO and NOx behavior (Figs. 1 and 4). The addition of PAN and N2O5 chemistry does not rectify the nighttime behavior of isoprene (Fig. 9) but do bring the cold-season-simulated CO and NOx mixing ratios closer to the MO mechanism (Fig. 8). These deficiencies may result from the highly parameterized biogenic chemistry within the SF mechanisms (Table S2), although it may also result from the treatment of isoprene emissions, and future simulations will need to consider the trade-off between additional complexity and computational efficiency.

More surprisingly, there are several desirable capabilities that are not lost with the selection of the SF mechanism. For most regions, the selection of the SF mechanism does not degrade the estimate of surface ozone (both the magnitude and the variability), nor do we lose features of the daily variability that results from the meteorology. In many regions, and at many scales, we find that the selection of the SF mechanism introduces uncertainties that are smaller than the difference between the simulated and observed surface ozone mixing ratios (Fig. 5). Surface layer ozone is adequately represented over many regions in all seasons within the SF mechanism (Fig. 8), despite the high CO and low NOx levels in the winter and spring seasons (Fig. 4). For these seasons, the adequate ozone representation may be the result of compensating errors, and Schnell et al. (2015) previously found comparable cases where the SF mechanism outperforms more complex models, perhaps due to various sets of compensating biases or errors.

We now turn to the main question of this research: what do we gain when we select a simplified chemical mechanism? The primary thing we gain is the capability to simulate longer periods of time or to include more members in an ensemble in proportion to the simplicity of the mechanism. Our results show that, without any optimization of the code, the RH mechanism is ∼100 % faster than the MO mechanism, and the SF mechanism is up to 200 % faster than the MO mechanism (Table 1). We feel that the capability to run three SF simulations for the price of one MO simulation under different sets of initial conditions, for example, can extend the quantification of parametric uncertainties, which is largely unavailable to the most complex and most computationally demanding mechanisms. Of course, the SF mechanism may not be appropriate for every sensitivity study but neither is the MO mechanism. The choice of mechanism really depends, then, on the science question. If the research objective is to predict complex chemistry–climate interactions and if computational resources are available, then a more complex mechanism will have the most value. However, if the research objective is to better understand various parameterizations, then a more computationally efficient mechanism will have higher value even if it might not be capable of accurately simulating all variables in detail (Hoffman et al., 1996). This is particularly the case when a baseline can be established between the simplified mechanism and the complex mechanism, as we have done here. We feel that this parallel approach, in which a set of mechanisms with varying levels of complexity are run concurrently with a consistent set of parameters, allows us to enhance our exploration of uncertainties and thus our ability to understand the atmospheric chemistry of the Earth system.

For instance, there are many research frameworks where the “three-for-one” advantage of the SF mechanism could be utilized with the MO mechanism, in which one simulation of a 5- or 10-year time slice with the MO mechanisms could be combined with three simulations of the SF mechanism, one matching the parameters of the MO mechanism (in order to provide a consistent baseline) and the other two exploring other parameter spaces (e.g., different initial conditions or different emission scenarios). The establishment of a baseline comparison is particularly important, since the SF mechanism is a simplified mechanism and should not be blindly trusted to reproduce the behavior of more complex mechanisms. For example, if a research group is interested in precise estimates of ozone concentrations in regions where the biogenic influence is significant, the SF mechanism would prove insufficient. The RH mechanism may be sufficient, but the more modest increase in computational speed – a “two-for-one” advantage over the MO mechanism – may not be enough to justify the simulation. If, however, the phenomenon of interest can be shown to be within the SF mechanism capabilities (e.g., simulating regional-scale ozone, as shown in this paper), the “three-for-one” advantage of the SF mechanism is readily apparent. The SF mechanism may be particularly desirable with chemistry–climate simulations at higher spatial resolutions.

In addition, the selection of a simplified mechanism allows for the capability to easily and efficiently test new forms and new representations of chemistry without the need to painstakingly update and test all possible interactions of any addition within a complex mechanism. For example, in this study, we added a simplified PAN and N2O5 representation to the SF mechanism (Figs. 8 and 9) to see how it improves the simulations. This exercise offered a significant capability to test, simulate, and further learn about improving atmospheric chemistry computations. This demonstrates that a hybrid approach (or tiered approach, as recommended in Uusitalo et al., 2015) – in which complex and trusted chemical mechanisms are used to evaluate simplified mechanisms that can run for longer periods or with increased ensemble members – has the potential to maximize computational capabilities and to get the most out of atmospheric chemistry modeling.

Furthermore, the selection of a simple chemical mechanism – especially when used in conjunction with more complex mechanisms within a consistent modeling framework – allows for better quantification of the uncertainties and the relative importance of particular pieces of the chemistry. Here, for instance, the SF mechanism's representation of biogenic species chemistry is insufficient to adequately represent equatorial landmasses, but the reduced-form RH mechanism is nearly as capable as the MO mechanism over most regions and most species. This begs the following question: is there a representation of biogenic chemistry somewhere between the RH and the SF mechanisms that can approach the efficiency of the SF mechanism and the accuracy of the RH mechanism? We hope that future research will address this question, as well as others, such as more globally oriented research pertaining to ozone budgets and the interaction between OH and CH4 lifetime. In addition, comparisons of chemical mechanisms of different complexities, particularly where the simplified mechanisms fail, could potentially identify regional chemical regimes. For instance, the SF mechanism cannot adequately represent the chemistry of equatorial forests (Fig. 7), and the spatial regions that fail to simulate ozone chemistry are similar to the spatial distribution of the tropical forest chemical regime identified in Fig. 4 of Sofen et al. (2016), which utilized a statistical clustering technique to identify chemical regimes. Finally, the capability to examine atmospheric chemistry complexity in a step-wise fashion could also be utilized to bridge the gap between the most complex 3-D chemical models and the more efficient models utilized by the Earth models of intermediate complexity (EMIC) or integrated assessment model (IAM) communities.

In this study, we have compared three chemical mechanisms of different levels of complexity within the CESM CAM-chem framework for present-day chemical and climatological conditions. We conducted 25-year cycled emission simulations nudged to MERRA meteorology with the standard tropospheric MOZART-4 (MO) mechanism of Emmons et al. (2010), the Reduced Hydrocarbon (RH) mechanism of Houweling et al. (1998), and the Super-Fast (SF) mechanism of Cameron-Smith et al. (2006). The RH mechanism is roughly twice as efficient as the MO mechanism, and the SF mechanism is roughly 3 times as efficient as the MO mechanism, without any code optimization. As much as possible, we kept the parameterizations consistent across all mechanisms, although we had to remap some of the MO mechanism species to match up with the RH mechanism species.

We examine present-day chemistry with MO, RH, and SF. Both MO and SF have been compared in other model intercomparisons, including for preindustrial conditions (see the Supplement for additional information). We hope that the analysis presented in this paper and the availability of the mechanism files (Supplement) will provide a baseline for continuing research on both the RH and SF mechanisms.

We find that all three mechanisms successfully capture surface ozone values at the larger spatial scales, but at smaller spatial scales, and especially within the northeastern US, all three mechanisms have surface ozone biases when compared to CASTNET observations; however, the mean values for all three mechanisms are consistent with each other at a variety of spatial scales. The SF mechanism simulations show larger ozone variability than the MO and RH simulations, although when normalizing the distributions to account for the known ozone biases, the SF mechanism represents the shape and spread of the ozone distributions better than the MO or RH mechanisms when compared to the CASTNET observations (Fig. 5).

The RH mechanism is in close agreement with the MO mechanism for nearly every metric we examined, and any differences tend to be minor (both in magnitude and in spatial extent). The SF mechanism simulates higher NOx and lower CO than the MO mechanism, and the NOx deviations are particularly large in the winter season. In addition, the SF mechanism deviates from the MO mechanism over regions of high biogenic emissions, such as equatorial Africa and South America. These large deviations within the SF mechanism are likely a result of the simplicity of the mechanism and especially of the lack of biogenic species chemistry beyond a single-species, two-reaction representation as well as a lack of PAN and N2O5 chemistry (Figs. 8 and 9). The SF mechanisms do not include NO3, which may also explain some of the nighttime biases. Future simulations in which NO3 chemistry is added to the SF mechanism may correct some of these biases. We also find that although the SF mechanism differs in the magnitude of the estimated ozone from the other two mechanisms, the simulated ozone variability is similar in all three mechanisms (Figs. 4 and 10).

We find that there are significant gains that can be realized by a research approach that utilizes simulations with both a complex and a simplified chemical mechanism where the complex mechanisms are used to provide a more trusted chemical result (especially for the mean values) and the simple mechanism could be used to efficiently simulate longer time periods to better understand the roles of meteorological variability. The capability of the SF mechanism to simulate adequate chemistry with interactive meteorology is not examined here nor is the coupling of the SF mechanism with modal aerosols, which is left for future research. These results encourage revitalizing or creating simplified chemical mechanisms within individual modeling frameworks and examining the structural uncertainties that exist between different models with regards to simplified chemical mechanisms.

Finally, we note that there are many inherent uncertainties associated with the use and comparison of chemical mechanisms and climate–chemistry simulations, many of which are inherited with the adoption of a particular model. The CESM CAM-chem model has been used extensively to examine a variety of climate and chemistry phenomena, and uncertainties that arise from the individual choices made during the historical development of this chemical model (see Brasseur et al., 1998; Hauglustaine et al., 1998; Horowitz et al., 2003; Kinnison et al., 2007; Emmons et al., 2010) are still present in the CESM CAM-chem modeling framework, such as which scheme or parameterization was to be included and the specific metric and methodology of tuning the climate model to historical data (see Hourdin et al., 2017, and references therein). Future simulations using different model versions or different choices of parameterizations, schemes, emissions, and other input datasets will need to examine the impact of those choices on the simulated chemical uncertainty and compare these to the uncertainty that arises from the selection of the different chemical mechanisms presented here.