Predicting secondary organic aerosol (SOA) formation relies either on extremely detailed, numerically expensive models accounting for the condensation of individual species or on extremely simplified, numerically affordable models parameterizing SOA formation for large-scale simulations. In this work, we explore the possibility of creating a random forest to reproduce the behavior of a detailed atmospheric organic chemistry model at a fraction of the numerical cost. A comprehensive dataset was created based on thousands of individual detailed simulations, randomly initialized to account for the variety of atmospheric chemical environments. Recurrent random forests were trained to predict organic matter formation from dodecane and toluene precursors, and the partitioning between gas and particle phases. Validation tests show that the random forests perform well without any divergence over 10 days of simulations. The distribution of errors shows that the sampling of initial conditions for the training simulations needs to focus on chemical regimes where SOA production is the most sensitive. Sensitivity tests show that specializing multiple random forests for a specific chemical regime is not more efficient than training a single general random forest for the entire dataset. The most important predictors are those providing information about the chemical regime, oxidants levels and existing organic mass. The choice of predictors is crucial as using too many unimportant predictors reduces the performances of the random forests.

Secondary organic aerosols (SOA) are formed from oxidation of hundreds of volatile organic compounds (VOCs) emitted from anthropogenic and natural sources. Accurate predictions of this chemistry are key for air quality and climate studies due to the large contribution of organic aerosols to submicron aerosol mass. Currently, only explicit models, such as the Generator for Explicit Chemistry and Kinetics of Organics in the Atmosphere (GECKO-A), can fully represent the chemical processing of thousands of organic species. However, their extreme computational cost prohibits their use in current chemistry-climate models, which rely on simplified empirical parameterizations to predict SOA concentrations. Recent applications of atmospheric chemistry emulation with machine learning (ML) applied to the simpler chemical mechanisms of tropospheric ozone have shown its ability to produce realistic predictions and significantly reduce the computational cost. This study proves that ML can accurately emulate SOA formation from an explicit chemistry model for several precursors with 100 to 100,000 times speedup over GECKO-A, making it computationally usable in a chemistry-climate model. To train the ML emulator, we generated thousands of GECKO-A box simulations sampled from a broad range of initial environmental conditions, and focused on the chemistry of three representative SOA precursors: the oxidation by OH of two anthropogenic (toluene, dodecane), and one biogenic VOC (alpha-pinene). We compare fully-connected and recurrent neural network methods and use an ensemble approach to quantify their underlying uncertainty and robustness. The SOA predictions generally remain stable over a simulation period of 5 days with an approximate error of 2-8\%.