Advancing Air Pollutant Prediction Through Deep Learning and Physics-Informed Machine Learning

Air pollution remains a critical challenge for public health and environmental sustainability, with PM_2.5. In this study, we develop a data-driven predictive framework based on a deep feedforward neural network to estimate PM_2.5 concentrations using meteorological data from two air quality monitoring stations in South East Queensland, Australia. Through exploratory data analysis, seven key meteorological variables were identified as input features, and the model demonstrated strong predictive performance in capturing overall temporal patterns of PM_2.5 variation. However, the model exhibited limitations in forecasting abrupt fluctuations, highlighting the constraints of purely data-driven approaches. To address these shortcomings, we propose the integration of physics-informed machine learning (PIML) methodologies, which embed domain-specific physical laws within the model architecture to enhance prediction robustness under conditions of noise and sparsity. This research contributes to the broader goal of fostering equitable and sustainable urban environments by advancing computational tools that can be deployed in data-scarce contexts. The findings intersect meaningfully with the lived experiences of communities exposed to elevated pollution levels and limited environmental oversight. By enabling accurate forecasting with minimal data requirements, the proposed approach holds the potential to inform policy decisions and support environmental justice advocacy. This work helps improve air quality modelling techniques while also ensuring that all communities can benefit from scientific advancements.