14 We present a methodology that uses gradient boosted regression trees (a machine learning 15 technique) and a full-chemistry simulation (i.e., training dataset) from a chemistry climate model (CCM) 16 to efficiently generate a parameterization of tropospheric hydroxyl radical (OH) that is a function of 17 chemical, dynamical, and solar irradiance variables. This surrogate model of OH is designed to allow for 18 computationally-efficient simulation of nonlinear feedbacks between OH and tropospheric constituents 19 that have loss by reaction with OH as their primary sinks (e.g., carbon monoxide (CO), methane (CH4), 20 volatile organic compounds (VOCs)). Such a model framework is advantageous for studies that require 21 multi-decadal simulations of CH4 or multi-year sensitivity simulations to understand the causes of trends 22 and variations of CO and CH4. The methodology that we present provides for the relatively easy creation 23 of a new parameterization in response to, for example, changes in the underlying CCM chemistry and/or 24 dynamics schemes. We show that a parameterization of OH generated from a CCM simulation is able to 25 reproduce OH concentrations with a normalized root mean square error of approximately 5%, as well as 26 capturing the global mean methane lifetime within approximately 1%. The accuracy of the 27 parameterization is dependent on inputs being within the bounds of the training dataset. However, we 28 show that the parameterization predicts large deviations in OH for an El Niño event that was not part of 29 the training dataset, and that the spatial distribution and strength of these deviations are consistent 30 with the event. This result gives confidence in the fidelity of the parameterization to simulate the spatial 31 and temporal responses of OH to perturbations from large variations in the chemical, dynamical and 32 solar irradiance drivers of OH. In addition, we discuss how two machine learning metrics, Gain feature 33 importance and SHAP values, indicate that the behavior of the parameterization of OH generally 34 comports with our understanding of OH chemistry, even though there are no physicsor chemistry35 based constraints on the parameterization. 36 37

We present a methodology that uses gradient-boosted regression trees (a machine learning technique) and a full-chemistry simulation (i.e., training dataset) from a chemistry–climate model (CCM) to efficiently generate a parameterization of tropospheric hydroxyl radical (OH) that is a function of chemical, dynamical, and solar irradiance variables. This surrogate model of OH is designed to be integrated into a CCM and allow for computationally efficient simulation of nonlinear feedbacks between OH and tropospheric constituents that have loss by reaction with OH as their primary sinks (e.g., carbon monoxide (CO), methane (CH4), volatile organic compounds (VOCs)). Such a model framework is advantageous for studies that require multi-decadal simulations of CH4 or multi-year sensitivity simulations to understand the causes of trends and variations of CO and CH4. To allow the user to easily target the training dataset towards a desired application, we are outlining a methodology to generate a parameterization of OH and not presenting an “off-the-shelf” version of a parameterization to be incorporated into a CCM. This provides for the relatively easy creation of a new parameterization in response to, for example, changes in research goals or the underlying CCM chemistry and/or dynamics schemes. We show that a sample parameterization of OH generated from a CCM simulation is able to reproduce OH concentrations with a normalized root-mean-square error of approximately 5 % and capture the global mean methane lifetime within approximately 1 %. Our calculated accuracy of the parameterization assumes inputs being within the bounds of the training dataset. Large excursions from these bounds will likely decrease the overall accuracy. However, we show that the sample parameterization predicts large deviations in OH for an El Niño event that was not part of the training dataset and that the spatial distribution and strength of these deviations are consistent with the event. This result gives confidence in the fidelity of a parameterization developed with our methodology to simulate the spatial and temporal responses of OH to perturbations from large variations in the chemical, dynamical, and solar irradiance drivers of OH. In addition, we discuss how two machine learning metrics, Gain feature importance and Shapley additive explanations values, indicate that the behavior of a parameterization of OH generally accords with our understanding of OH chemistry, even though there are no physics- or chemistry-based constraints on the parameterization.

Atmospheric ozone (O 3 ) is a pollutant produced through chemical chain reactions where volatile organic compounds (VOCs), carbon monoxide and methane are oxidized in the presence of oxides of nitrogen (NO x ). For decades, the controlling chain termination step has been used to separate regions into either ‘NO x limited’ (peroxyl-radical self-reactions dominate) or ‘VOC limited’ (hydroxyl radical (OH) + nitrogen dioxide (NO 2 ) reaction dominates). The controlling regime would then guide policies for reducing emissions and so O 3 concentrations. Using a chemical transport model, we show that a third ‘aerosol inhibited’ regime exists, where reactive uptake of hydroperoxyl radicals (HO 2 ) onto aerosol particles dominates. In 1970, 2% of the Northern Hemisphere population lived in an aerosol-inhibited regime, but by 2014 this had increased to 21%; 60% more than lived in a VOC-limited regime. Aerosol-inhibited chemistry suppressed surface O 3 concentrations in North America and Europe in the 1970s and is currently suppressing surface O 3 over Asia. This third photochemical O 3 regime leads to potential trade-off tensions between reducing particle pollution in Asia (a key current health policy and priority) and increasing surface O 3 , should O 3 precursors emissions not be reduced in tandem.

A sample parameterization of OH for July created to be used with the ECCOH module of the GEOS Earth System Model. The methodology to generate the parameterization is described in detail in the GMD article "A Machine Learning Methodology for the Generation of a Parameterization of the Hydroxyl Radical". The parameterization is included as an example and should not be used for research purposes without consulting the publication and/or the authors. We have also included the dataset used to train the parameterization (OHParameterization_TrainingSet_1980_2019_M07.dat) as well as the training targets (OHParameterization_Targets_1980_2019_M07.dat). The training data are from the NASA MERRA2 GMI simulation (https://acd-ext.gsfc.nasa.gov/Projects/GEOSCCM/MERRA2GMI/). The scripts used to generate the training dataset and the parameterization can be found at https://zenodo.org/record/6046037.

Abstract. We present a methodology that uses gradient boosted regression trees (a machine learning technique) and a full-chemistry simulation (i.e., training dataset) from a chemistry climate model (CCM) to efficiently generate a parameterization of tropospheric hydroxyl radical (OH) that is a function of chemical, dynamical, and solar irradiance variables. This surrogate model of OH is designed to allow for computationally-efficient simulation of nonlinear feedbacks between OH and tropospheric constituents that have loss by reaction with OH as their primary sinks (e.g., carbon monoxide (CO), methane (CH4), volatile organic compounds (VOCs)). Such a model framework is advantageous for studies that require multi-decadal simulations of CH4 or multi-year sensitivity simulations to understand the causes of trends and variations of CO and CH4. The methodology that we present provides for the relatively easy creation of a new parameterization in response to, for example, changes in the underlying CCM chemistry and/or dynamics schemes. We show that a parameterization of OH generated from a CCM simulation is able to reproduce OH concentrations with a normalized root mean square error of approximately 5 %, as well as capturing the global mean methane lifetime within approximately 1 %. The accuracy of the parameterization is dependent on inputs being within the bounds of the training dataset. However, we show that the parameterization predicts large deviations in OH for an El Niño event that was not part of the training dataset, and that the spatial distribution and strength of these deviations are consistent with the event. This result gives confidence in the fidelity of the parameterization to simulate the spatial and temporal responses of OH to perturbations from large variations in the chemical, dynamical and solar irradiance drivers of OH. In addition, we discuss how two machine learning metrics, Gain feature importance and SHAP values, indicate that the behavior of the parameterization of OH generally comports with our understanding of OH chemistry, even though there are no physics- or chemistry-based constraints on the parameterization.

Aromatic hydrocarbons, including benzene, toluene, and xylenes, play an important role in atmospheric chemistry, but the associated chemical mechanisms are complex and uncertain. Sparing representation of this chemistry in models is needed for computational tractability. Here, we develop a new compact mechanism for aromatic chemistry (GC13) that captures current knowledge from laboratory and computational studies with only 17 unique species and 44 reactions. We compare GC13 to six other currently used mechanisms of varying complexity in box model simulations of environmental chamber data and diurnal boundary layer chemistry, and show that GC13 provides results consistent with or better than more complex mechanisms for oxygenated products (alcohols, carbonyls, dicarbonyls), ozone, and hydrogen oxide (HOx≡OH+HO2) radicals. Specifically, GC13 features increased radical recycling and increased ozone destruction from phenoxy–phenylperoxy radical cycling relative to other mechanisms. We implement GC13 into the GEOS-Chem global chemical transport model and find higher glyoxal yields and net ozone loss from aromatic chemistry compared with other mechanisms. Aromatic oxidation in the model contributes 23 %, 5 %, and 8 % of global glyoxal, methylglyoxal, and formic acid production, respectively, and has mixed effects on formaldehyde. It drives small decreases in global tropospheric OH (−2.2 %), NOx (≡NO+NO2; −3.7 %), and ozone (−0.8 %), but a large increase in NO3 (+22 %) from phenoxy–phenylperoxy radical cycling. Regional effects in polluted environments can be substantially larger, especially from the photolysis of carbonyls produced by aromatic oxidation, which drives large wintertime increases in OH and ozone concentrations.

Effective mitigation of surface ozone pollution entails detailed knowledge of the contributing precursors' sources. We use the GEOS-Chem adjoint model to analyze the precursors contributing to surface ozone in the Beijing-Tianjin-Hebei area (BTH) of China on days of different ozone pollution severities in June 2019. We find that BTH ozone on heavily polluted days is sensitive to local emissions, as well as to precursors emitted from the provinces south of BTH (Shandong, Henan, and Jiangsu, collectively the SHJ area). Heavy ozone pollution in BTH can be mitigated effectively by reducing NOx (from industrial processes and transportation), ≥C3 alkenes (from on-road gasoline vehicles and industrial processes), and xylenes (from paint use) emitted from both BTH and SHJ, as well as by reducing CO (from industrial processes, transportation, and power generation) and ≥C4 alkanes (from industrial processes, paint and solvent use, and on-road gasoline vehicles) emissions from SHJ. In addition, reduction of NOx, xylene, and ≥C3 alkene emissions within BTH would effectively decrease the number of BTH ozone-exceedance days. Our analysis pinpoint the key areas and activities for locally and regionally coordinated emission control efforts to improve surface ozone air quality in BTH.

In assessments of cancer risk from atmospheric polycyclic aromatic hydrocarbons (PAHs), scientists and regulators rarely consider the complex mixture of emitted compounds and degradation products, and they often represent the entire mixture using a single emitted compound-benzo[a]pyrene. Here, we show that benzo[a]pyrene is a poor indicator of PAH risk distribution and management: nearly 90% of cancer risk worldwide results from other PAHs, including unregulated degradation products of emitted PAHs. We develop and apply a global-scale atmospheric model and conduct health impact analyses to estimate human cancer risk from 16 PAHs and several of their N-PAH degradation products. We find that benzo[a]pyrene is a minor contributor to the total cancer risks of PAHs (11%); the remaining risk comes from other directly emitted PAHs (72%) and N-PAHs (17%). We show that assessment and policy-making that relies solely on benzo[a]pyrene exposure provides misleading estimates of risk distribution, the importance of chemical processes, and the prospects for risk mitigation. We conclude that researchers and decision-makers should consider additional PAHs as well as degradation products.

The impact of emissions of volatile organic compounds (VOCs) to the atmosphere on the production of secondary pollutants, such as ozone and secondary organic aerosol (SOA), is mediated by the concentration of nitric oxide (NO). Polluted urban atmospheres are typically considered to be “high-NO” environments, while remote regions such as rainforests, with minimal anthropogenic influences, are considered to be “low NO”. However, our observations from central Beijing show that this simplistic separation of regimes is flawed. Despite being in one of the largest megacities in the world, we observe formation of gas- and aerosol-phase oxidation products usually associated with low-NO “rainforest-like” atmospheric oxidation pathways during the afternoon, caused by extreme suppression of NO concentrations at this time. Box model calculations suggest that during the morning high-NO chemistry predominates (95 %) but in the afternoon low-NO chemistry plays a greater role (30 %). Current emissions inventories are applied in the GEOS-Chem model which shows that such models, when run at the regional scale, fail to accurately predict such an extreme diurnal cycle in the NO concentration. With increasing global emphasis on reducing air pollution, it is crucial for the modelling tools used to develop urban air quality policy to be able to accurately represent such extreme diurnal variations in NO to accurately predict the formation of pollutants such as SOA and ozone.

Predictions from process-based models of environmental systems are biased, due to uncertainties in their inputs and parameterizations, reducing their utility. We develop a predictor for the bias in tropospheric ozone (O3, a key pollutant) calculated by an atmospheric chemistry transport model (GEOS-Chem), based on outputs from the model and observations of ozone from both the surface (EPA, EMEP, and GAW) and the ozone-sonde networks. We train a gradient-boosted decision tree algorithm (XGBoost) to predict model bias (model divided by observation), with model and observational data for 2010–2015, and then we test the approach using the years 2016–2017. We show that the bias-corrected model performs considerably better than the uncorrected model. The root-mean-square error is reduced from 16.2 to 7.5 ppb, the normalized mean bias is reduced from 0.28 to −0.04, and Pearson's R is increased from 0.48 to 0.84. Comparisons with observations from the NASA ATom flights (which were not included in the training) also show improvements but to a smaller extent, reducing the root-mean-square error (RMSE) from 12.1 to 10.5 ppb, reducing the normalized mean bias (NMB) from 0.08 to 0.06, and increasing Pearson's R from 0.76 to 0.79. We attribute the smaller improvements to the lack of routine observational constraints for much of the remote troposphere. We show that the method is robust to variations in the volume of training data, with approximately a year of data needed to produce useful performance. Data denial experiments (removing observational sites from the algorithm training) show that information from one location (for example Europe) can reduce the model bias over other locations (for example North America) which might provide insights into the processes controlling the model bias. We explore the choice of predictor (bias prediction versus direct prediction) and conclude both may have utility. We conclude that combining machine learning approaches with process-based models may provide a useful tool for improving these models.

Abstract. The impact of volatile organic compound (VOC) emissions to the atmosphere on the production of secondary pollutants, such as ozone and secondary organic aerosol (SOA), is mediated by the concentration of nitric oxide (NO). Polluted urban atmospheres are typically considered to be high-NO environments, while remote regions such as rainforests, with minimal anthropogenic influences, are considered to be low-NO. Policy to reduce urban air pollution is typically developed assuming that the chemistry is controlled by the high-NO regime. However, our observations from central Beijing show that this simplistic separation of regimes is flawed. Despite being in one of the largest megacities in the world, we observe significant formation of gas and aerosol phase oxidation products associated with the low-NO rainforest-like regime during the afternoon. This is caused by a surprisingly low concentration of NO, coupled with high concentrations of VOCs and of the atmospheric oxidant hydroxyl (OH). Box model calculations suggest that during the morning high-NO chemistry predominates (95 %) but in the afternoon low-NO chemistry plays a greater role (30 %). With increasing global emphasis on reducing air pollution, the modelling tools used to develop urban air quality policy need to adequately represent both high- and low-NO regimes if they are to have utility.

Low-cost sensors (LCSs) are an appealing solution to the problem of spatial resolution in air quality measurement, but they currently do not have the same analytical performance as regulatory reference methods. Individual sensors can be susceptible to analytical cross-interferences; have random signal variability; and experience drift over short, medium and long timescales. To overcome some of the performance limitations of individual sensors we use a clustering approach using the instantaneous median signal from six identical electrochemical sensors to minimize the randomized drifts and inter-sensor differences. We report here on a low-power analytical device (< 200 W) that is comprised of clusters of sensors for NO2, O-x, CO and total volatile organic compounds (VOCs) and that measures supporting parameters such as water vapour and temperature. This was tested in the field against reference monitors, collecting ambient air pollution data in Beijing, China. Comparisons were made of NO2 and O-x clustered sensor data against reference methods for calibrations derived from factory settings, in-field simple linear regression (SLR) and then against three machine learning (ML) algorithms. The parametric supervised ML algorithms, boosted regression trees (BRTs) and boosted linear regression (BLR), and the non-parametric technique, Gaussian process (GP), used all available sensor data to improve the measurement estimate of NO2 and O-x. In all cases ML produced an observational value that was closer to reference measurements than SLR alone. In combination, sensor clustering and ML generated sensor data of a quality that was close to that of regulatory measurements (using the RMSE metric) yet retained a very substantial cost and power advantage.