The Cloud Condensation Nuclei (CCN) activity of aerosols remains one of the major challenges in understanding the aerosol-cloud interaction process. Not only particle size and composition, but also environmental factors and atmospheric dynamics influence CCN activity. A comprehensive machine learning framework was adopted to identify the factors influencing CCN activity over a rural location Gadanki, India. The dependency of CCN on 18 variables (aerosol, precursor gas, and meteorological parameters) was quantified using three shrinkage methods (Ridge, Lasso, and Elastic Net), two dimensionality reduction methods (Principal Component Regression, PCR; Partial Least Square Regression, PLSR), and three machine learning methods (Random Forest, RF; Support Vector Machines, SVM; Artificial Neural Networks, ANN). Temperature exhibited the highest multicollinearity among all variables, followed by relative humidity (RH), boundary layer height, and ozone (O3). Model performance was evaluated by randomly selecting 70% of the common dataset for training and 30% for testing, at a seasonal scale and separately for day and night. The evaluation metrics showed that the testing R2 was highest (> 60%) for RF and lowest (< 30%) for PCR, indicating that RF is the best-suited model. The factors influencing CCN, as estimated using Ridge, Lasso, Elastic Net, RF, and SVM, showed strong consensus across models. The O3 and the absorption coefficient prominently influence CCN during the pre-monsoon (Mar-May) and winter (Dec-Feb) seasons, while CO, RH and O3 were influential during the monsoon (Jun-Sep). Similarly, the scattering coefficient and particle concentration (> 0.5 micron) were influential during the post-monsoon (Oct-Nov) season. The results presented here reiterate the key roles played by biomass burning, chemical transformations, and ozone oxidation. It is concluded that, alongside physics-based models, machine learning approaches are valuable tools for understanding CCN activity. Graphical Abstract The graphical abstract provides a summary of work carried to identify factors influencing the Cloud Condensation Nuclei (CCN) activity of aerosols over a rural location, Gadanki in India using Machine Learning framework. CCN is the fraction of atmospheric aerosols that nucleates the cloud droplets. The CCN activity of aerosols not only depends on the aerosol properties such as their size, and chemical composition, but also on the background environment, and dynamics. A comprehensive machine learning based framework has been adopted using shrinkage methods, dimensionality reduction methods, and machine learning techniques. A total of 18 parameters comprising of aerosol (Na, CN, σa, σs), gases (CO, O3, SO2, NO2, NH3, VOCs), and meteorological (T, RH, WS, WD, BLH, SR) information obtained using collocated instruments have been provided as an input to the machine learning framework considering CCN as dependent variable and all other parameters as independent variables. The evaluation metrics of models suggests, Random Forest as the best suitable model with highest R2 > 60%. The factors influencing the CCN were identified both for day and night separately at seasonal scales. Variable importance shows O3, CO, SO2, NH3, σa, and σs have the highest influence among different seasons. The results emphasize the role of bio-mass burning, ozone oxidation, and photo-chemical transformation in CCN activity over a rural location.

A Machine Learning Framework for Factors Influencing Cloud Condensation Nuclei

期刊论文
发表日期：

2025年11月1日

查看全文

V. Ravi Kiran,
Tamada Jayasri,
M. Venkat Ratnam

摘要

The Cloud Condensation Nuclei (CCN) activity of aerosols remains one of the major challenges in understanding the aerosol-cloud interaction process. Not only particle size and composition, but also environmental factors and atmospheric dynamics influence CCN activity. A comprehensive machine learning framework was adopted to identify the factors influencing CCN activity over a rural location Gadanki, India. The dependency of CCN on 18 variables (aerosol, precursor gas, and meteorological parameters) was quantified using three shrinkage methods (Ridge, Lasso, and Elastic Net), two dimensionality reduction methods (Principal Component Regression, PCR; Partial Least Square Regression, PLSR), and three machine learning methods (Random Forest, RF; Support Vector Machines, SVM; Artificial Neural Networks, ANN). Temperature exhibited the highest multicollinearity among all variables, followed by relative humidity (RH), boundary layer height, and ozone (O₃). Model performance was evaluated by randomly selecting 70% of the common dataset for training and 30% for testing, at a seasonal scale and separately for day and night. The evaluation metrics showed that the testing R² was highest (> 60%) for RF and lowest (< 30%) for PCR, indicating that RF is the best-suited model. The factors influencing CCN, as estimated using Ridge, Lasso, Elastic Net, RF, and SVM, showed strong consensus across models. The O₃ and the absorption coefficient prominently influence CCN during the pre-monsoon (Mar-May) and winter (Dec-Feb) seasons, while CO, RH and O₃ were influential during the monsoon (Jun-Sep). Similarly, the scattering coefficient and particle concentration (> 0.5 micron) were influential during the post-monsoon (Oct-Nov) season. The results presented here reiterate the key roles played by biomass burning, chemical transformations, and ozone oxidation. It is concluded that, alongside physics-based models, machine learning approaches are valuable tools for understanding CCN activity.

Graphical Abstract

The graphical abstract provides a summary of work carried to identify factors influencing the Cloud Condensation Nuclei (CCN) activity of aerosols over a rural location, Gadanki in India using Machine Learning framework. CCN is the fraction of atmospheric aerosols that nucleates the cloud droplets. The CCN activity of aerosols not only depends on the aerosol properties such as their size, and chemical composition, but also on the background environment, and dynamics. A comprehensive machine learning based framework has been adopted using shrinkage methods, dimensionality reduction methods, and machine learning techniques. A total of 18 parameters comprising of aerosol (N_a, CN, σ_a, σ_s), gases (CO, O₃, SO₂, NO₂, NH₃, VOCs), and meteorological (T, RH, WS, WD, BLH, SR) information obtained using collocated instruments have been provided as an input to the machine learning framework considering CCN as dependent variable and all other parameters as independent variables. The evaluation metrics of models suggests, Random Forest as the best suitable model with highest R² > 60%. The factors influencing the CCN were identified both for day and night separately at seasonal scales. Variable importance shows O₃, CO, SO₂, NH₃, σ_a, and σ_s have the highest influence among different seasons. The results emphasize the role of bio-mass burning, ozone oxidation, and photo-chemical transformation in CCN activity over a rural location.