Predicting Air Pollutant Concentration Using Deep Neural Network: Path towards a Physics-Informed Machine Learning Framework

Air pollution poses significant impacts to both human health and the global climate, requiring robust predictive models for accurate simulation of pollutant dispersion. The concentration of air pollutants exhibits considerable variability across short spatial distances due to diverse emission sources, meteorological conditions, and chemical processes. Traditional models often struggle to capture these complex and nonlinear atmospheric scenarios, leading to inaccuracies in their predictions; to overcome these limitations Machine Learning-particularly Physics Informed Machine Learning offers a promising solution. The primary objective of this study is to develop a purely data driven deep neural network model to capture pollutant dispersion without incorporating explicit physical knowledge. In this study, two monitoring stations in South East Queensland were selected and a deep neural network model was trained selecting PM_2.5 as the pollutant. A preliminary data analysis is conducted to identify the key meteorological factors that directly influence PM_2.5 concentration. Through this analysis, seven meteorological parameters were identified, which were used as inputs for feedforward neural network to predict future PM_2.5 concentrations. The accuracy of the model is evaluated using performance metrics and the results reveal that the model is closely capturing underlying patterns. However, the study exhibits noticeable discrepancies in predicting rapid concentration changes. To overcome such limitations, Physics Informed Machine Learning has emerged as a significant research avenue, as it is capable of making accurate predictions even under noisy and sparse data conditions