Machine Learning Reveals the Parameters Affecting the Gaseous Sulfuric Acid Distribution in a Coastal City: Model Construction and Interpretation

Abstract

Although the fundamental mechanisms of atmospheric new particle formation events are largely associated with gaseous sulfuric acid monomer (SA), the parameters affecting SA generation and elimination remain unclear, especially in coastal areas where certain sulfur-containing precursors are abundant. In this study, we utilized machine learning (ML) in combination with field observations to map the link between SA and the influencing parameters. The developed random forest (RF) model performed well in creating simulations with an R 2 of 0.90, and the significant factors were ultraviolet, methanesulfonic acid (MSA), SO2, condensation sink, and relative humidity in descending order. Among the five factors, MSA served as an indicator for sulfur-containing species from marine emissions. The black box of ML was broken to determine the marginal contribution of these five parameters to the model output using partial dependence plots and centered-individual conditional expectation plots. These results indicated that MSA had a positive impact on the performance of the RF model, and a co-occurring relationship was observed between MSA and SA during the nocturnal period. Our findings reveal that sulfur-containing species emitted from the marine environment have an impact on the formation of SA and should be considered in coastal areas.

New Insights on the Formation of Nucleation Mode Particles in a Coastal City Based on a Machine Learning Approach

“…A scanning mobility particle sizer based on electrical mobility size classification and an aerodynamic particle sizer based on the aerodynamic diameter were deployed to achieve a comprehensive measurement of particles (7–15200 nm). The atmospheric concentrations of methanesulfonic acid (CH 4 SO 3 ), LOOMs (log 10 C * ≤ – 0.5, where C * is the saturated vapor concentration), iodic acid (HIO 3 ), methyl amine (CH 3 NH 2 ), diethyl amine (C 4 H 9 NH 2 ), and sulfuric acid (H 2 SO 4 ) were determined using a chemical ionization time-of-flight mass spectrometer equipped with a nitrate source (nitrate-CIMS), as described in detail in our previous publication . Refer to Text S1 of the Supporting Information for in-depth insight into instrument operations, species identification, and uncertainty deliberations.…”

“…The atmospheric concentrations of methanesulfonic acid (CH 4 SO 3 ), LOOMs (log 10 C* ≤ − 0.5, where C* is the saturated vapor concentration), iodic acid (HIO 3 ), methyl amine (CH 3 NH 2 ), diethyl amine (C 4 H 9 NH 2 ), and sulfuric acid (H 2 SO 4 ) were determined using a chemical ionization time-of-flight mass spectrometer equipped with a nitrate source (nitrate-CIMS), as described in detail in our previous publication. 45 Refer to Text S1 of the Supporting Information for in-depth insight into instrument operations, species identification, and uncertainty deliberations. Supplementary details concerning ancillary data sources can be found in Text S2 of the Supporting Information.…”

“…41−44 For instance, our research team utilized a similar ML approach with data from the same site to develop a model for simulating gaseous H 2 SO 4 concentrations and assess the influence of different feature parameters on the output variable. 45 ML techniques adeptly capture intricate nonlinear relationships, paving the way for comprehensive interpretations. However, the complex dynamics of PNCN and the scarcity of systematic data sets present challenges when applying ML to the simulation of ultrafine particle number concentrations, particularly in relation to NMP.…”

“…While the studies highlighted earlier illuminate the PNCN determinants under diverse environmental contexts, identifying the key parameters that affect PNCN and investigating the marginal effects of individual (or multiple) feature variables on PNCN remain challenging. Machine learning (ML) has the potential to handle and learn from large, complex, and multidimensional data sets to develop simulative models. , ML methods, such as random forest (RF), support vector machine (SVM), and artificial neural network (ANN), have been applied to model pollutants in the atmosphere. − For instance, our research team utilized a similar ML approach with data from the same site to develop a model for simulating gaseous H 2 SO 4 concentrations and assess the influence of different feature parameters on the output variable . ML techniques adeptly capture intricate nonlinear relationships, paving the way for comprehensive interpretations.…”

Identifying Driving Factors of Atmospheric N2O5 with Machine Learning

“…Recently, machine learning has garnered significant attention in environmental science. − Utilizing data-driven models like tree models, machine learning excels in revealing intricate and concealed nonlinear correlations that traditional analytical methods often overlook. , However, machine learning algorithms are often regarded as a “black box” that makes it difficult to explain the underlying physical and chemical mechanisms. Fortunately, post-hoc interpretable models such as Shapley Additive exPlanations (SHAP) ,,− have been introduced, alleviating the model’s intricacy and improving its interpretability.…”