<p>Urban air quality in megacities such as Delhi is strongly modulated by meteorological variability, yet these interactions remain insufficiently represented in predictive modelling frameworks. This study evaluates the influence of key atmospheric parameters on PM2.5 concentrations using a benchmarking approach that integrates geo-spatial datasets with linear and non-linear machine learning models. Monthly observations from two representative monitoring locations, an industrial site (Bawana) and an urban-residential site (Okhla), selected through a network-wide parametric assessment were analysed to capture spatially distinct pollution-meteorology relationships. Multiple Linear Regression (MLR) was employed as a baseline, while a suite of Artificial Neural Network (ANN) architectures with varying complexity and regularization strategies were developed and optimized through systematic hyperparameter tuning; like the standard three-layer model; two-hidden layers; reduced-complexity with dropout regularization; and with batch normalization. Model evaluation indicated that ANN-based approaches consistently outperform linear regression, achieving higher explanatory power (R² generally &gt; 0.85) and lower prediction errors (RMSE typically reduced by ~ 15–25%) across both Bawana and Okhla stations. These improvements reflect the ability of ANN models to capture non-linear processes governing pollutant dispersion, secondary aerosol formation, and meteorological modulation, particularly in dense urban environments. Site-dependent variations further emphasize the influence of localized emissions and micro-meteorology on the predictive skill. Overall, the findings demonstrate that tailored deep learning frameworks provide a robust and scalable approach for PM2.5 prediction and can meaningfully support anticipatory air-quality management and evidence-based policy interventions in Delhi.</p>