Prediction of heavy metals in airborne fine particulate matter using magnetic parameters by machine learning from a metropolitan city in China

Abstract

This study explored a new statistical method for predicting airborne heavy metals using environmental magnetism. Samples of fine particulate matter (PM2.5) were collected for 1 year in a typical Chinese megacity. The annual average PM2.5 concentration was 60.17 μg/m3 and approximately 32.02% of daily PM2.5 concentrations exceeded the Chinese National Ambient Air Quality Standard guideline value. Concentrations of most heavy metals were higher in winter and lower in summer. The hazard index value, which reflects the health risk associated with exposure to potentially toxic metals via inhalation, was 0.506 and below the safe limit. The integrated carcinogenic risk was 1.64 × 10−5 for adults and 4.10 × 10−6 for children. Magnetic analysis indicated that soft low-coercivity ferrimagnetic minerals were the main magnetic material which pseudo single domain, or a mixture of single domain and multidomain magnetite dominated in PM2.5. Metal concentrations were simulated by using atmospheric pollutants, meteorological variables or magnetic parameters by support vector machine or backpropagation artificial neural network models; the former model outperformed the latter. Predictions for As, Cd, Pb, and Mn were better than for other metals with training and validation R values both >0.8, whereas relatively poor for Al, Co, and Cr. When magnetic parameters were included in the models, the prediction performance for all metals was improved; however, greater improvement was obtained for Cu, Cr, Fe, Ti, and V. Environmental magnetism analyses combined with machine learning have potential for efficient evaluation of airborne heavy metals.