Beyond Binary Maps Using HCHO/NO2 Ratios: First Estimates of Ozone Production Rates Along With Spatially-Varying Ozone Sensitivity Maps Using Satellite Retrievals in Conjunction with Deep Neural Networks

Our current knowledge about the global spatiotemporal variations of PO₃ is currently limited to chemical transport models, which can have significant uncertainties. However, we show a new method for parameterizing PO₃ using historical aircraft observations and a set of variables that can be largely informed by satellite observations (such as TROPOMI and OMI). One major advantage of this new approach, which uses deep neural networks, is that it eliminates the need to depend on HCHO/NO₂ (FNR) observations. FNR values can possess large errors due to chemical interactions between NO₂ and HCHO, as well as unresolved systematic and random errors in satellite retrievals that can be exacerbated in the ratio. Additionally, it is challenging to establish the right thresholds for distinguishing the continuously changing sensitivity of PO₃ to its precursors, as these thresholds can vary among different locations. We demonstrate an algorithm that utilizes well-constrained box models based on several million data points generated from aircraft datasets (both real and synthetic) under various photolysis rates as well as variations in meteorological and chemical conditions. This approach uses a regularized deep neural network algorithm, which can readily provide the contributions of NO₂ (as a surrogate for NO_x) and HCHO (as a surrogate for VOC reactivity) with much more spatial information compared to binary maps obtained from FNRs.