Improving the monoterpene oxidation scheme in a global-scale model through neural network-based bias correction

Monoterpenes emitted by vegetation, among other biogenic volatile organic compounds (BVOC), can play an important role in the formation of secondary organic aerosol (SOA). In reactions with atmospheric oxidants, monoterpenes can form products that condense into SOA. As the emissions of monoterpenes are temperature-dependent, a climate feedback is formed where rising temperatures increase biogenic SOA formation, which then cools the climate through aerosol-radiation interactions.

The accurate representation of this feedback mechanism would be important when modeling future climate. This necessitates a model of the underlying monoterpene oxidation chemistry to account for the volatilities of the oxidation products in varying conditions. Such models have been developed in recent years, one of which is the chamber chemistry model ADCHAM. ADCHAM is extended by the Peroxy Radical Autoxidation Mechanism to include monoterpene oxidation pathways, particularly for α-pinene. For climate applications, ADCHAM remains too complex without heavy simplification.

Our study aims to produce a parametrization of ADCHAM capable of predicting the volatility distribution of α-pinene oxidation products in simulations of the global atmosphere. To this end, we have trained a neural network (NN) to model the error between the current parametrization in the SALSA aerosol model and the more accurate ADCHAM in various conditions. We represent these conditions by eight input variables, including temperature and oxidant concentrations. The training data was generated by sampling points from the atmospheric ranges of the input variables in global reanalysis and climate model datasets, a subset of which were reserved for testing. For each point, ADCHAM was run for 7.5 minutes, corresponding to the timestep of our targeted climate model. The resulting compounds were aggregated into three bins based on their volatilities, according to the volatility basis set (VBS) representation used in SALSA. The differences between the VBS bin production rates (1/cm³s) from SALSA and ADCHAM constitute the training targets of the NN. For comparison, a linear regression model was also fitted.

We have tested the NN and linear model on the holdout set and found both to be successful in correcting the VBS concentrations produced by SALSA to match those from ADCHAM. Without correction, the SALSA representation generally resulted in higher production rates of the VBS bins compared to ADCHAM, in some cases by more than ten orders of magnitude (RMSE=5.03, i.e., five orders of magnitude). While the linear model corrects the overestimation and improves the fit (RMSE=1.97), errors as large as five orders of magnitude remain. Using the NN, such errors are eliminated – the NN-augmented SALSA corresponds remarkably well to ADCHAM (RMSE=0.28; R2=0.995). Additionally, the NN improved the modeled dependences between input variables and VBS bin production. The results are encouraging, suggesting that the dependence of condensable vapor production on ambient conditions in global models could be represented by augmenting simplified VBS schemes already in use with NNs.  While our NN is relatively small, further pruning seems possible without significantly affecting its accuracy. Before implementing the correction scheme into a global model, we will evaluate its ability to reproduce SOA yields in chamber simulations.