Atmospheric chemistry models-components in models that simulate air pollution and climate change-are computationally expensive. Previous studies have shown that machine-learned atmospheric chemical solvers can be orders of magnitude faster than traditional integration methods but tend to suffer from numerical instability. Here, we present a modeling framework that reduces error accumulation compared to previous work while maintaining computational efficiency. Our approach is novel in that it (1) uses a recurrent training regime that results in extended (>1 week) simulations without exponential error accumulation and (2) can reversibly compress the number of modeled chemical species by >80% without further decreasing accuracy. We observe an~260× speedup (~1,900× with specialized hardware) compared to the traditional solver. We use random initial conditions in training to promote general applicability across a wide range of atmospheric conditions. For ozone (concentrations ranging from 0-70 ppb), our model predictions over a 24-hr simulation period match those of the reference solver with median error of 2.7 and <19 ppb error across 99% of simulations initialized with random noise. Error can be significantly higher in the remaining 1% of simulations, which include extreme concentration fluctuations simulated by the reference model. Results are similar for total particulate matter (median error of 16 and <32 μg/m 3 across 99% of simulations with concentrations ranging from 0-150 μg/m 3). Finally, we discuss practical implications of our modeling framework and next steps for improvements. The machine learning models described here are not yet replacements for traditional chemistry solvers but represent a step toward that goal.

Acetone is one of the most abundant carbonyl compounds in the atmosphere, and it serves as an important source of HOx (OH + HO2) radicals in the upper troposphere and a precursor for peroxyacetyl nitrate. We present a global sensitivity analysis targeted at several major natural source and sink terms in the global acetone budget to find the input factor or factors to which the simulated acetone mixing ratio was most sensitive. The ranges of input factors were taken from literature. We calculated the influence of these factors in terms of their elementary effects on model output. Of the six factors tested here, the four factors with the highest contribution to total global annual model sensitivity are direct emissions of acetone from the terrestrial biosphere, acetone loss to photolysis, the concentration of acetone in the ocean mixed layer, and the dry deposition of acetone to ice‐free land. The direct emissions of acetone from the terrestrial biosphere are globally important in determining acetone mixing ratios, but their importance varies seasonally outside the tropics. Photolysis is most influential in the upper troposphere. Additionally, the influence of the oceanic mixed layer concentrations are relatively invariant between seasons, compared to the other factors tested. Monoterpene oxidation in the troposphere, despite the significant uncertainties in acetone yield in this process, is responsible for only a small amount of model uncertainty in the budget analysis.

Considerable financial resources are allocated for measuring ambient air pollution in the United States, yet the locations for these monitoring sites may not be optimized to capture the full extent of current pollution variability. Prior research on best sensor placement for monitoring fine particulate matter (PM2.5) pollution is scarce: most studies do not span areas larger than a medium-sized city or examine timescales longer than one week. Here we present a pilot study using multiresolution modal decomposition (mrDMD) to identify the optimal placement of PM2.5 sensors from 2000-2016 over the contiguous United States. This novel approach incorporates the variation of PM2.5 on timescales ranging from one day to over a decade to capture air pollution variability. We find that the mrDMD algorithm identifies more high-priority sensor locations in the western United States than those expected along the eastern coast, where a large number of EPA PM2.5 monitors currently reside. Specifically, 53% of mrDMD optimized sensor locations are west of the 100th meridian, compared to only 32% in the current EPA network. The mrDMD sensor locations can capture PM2.5 from wildfires and high pollution events, with particularly high skill in the West. These results suggest significant gaps in the current EPA monitoring network in the San Joaquin Valley in California, northern California, and in the Pacific Northwest (Idaho, and Eastern Washington and Oregon). Our framework diagnoses where to place air quality sensors so that they can best monitor smoke from wildfires. Our framework may also be applied to urban areas for equitable placement of PM2.5 monitors.