Atmospheric dispersion models are used to estimate the air quality impact of pollutants emitted by industrial plants. Their high computational demand hinders the development of real-time solutions that anticipate the effects of incidental emissions (estimating immissions) or infer the causes of abnormal pollution levels (estimating emissions). To address this limitation, this study explores the use of machine learning (ML) algorithms as surrogate models, aiming to replicate the results of dispersion models while reducing computational cost. The study evaluates different ML models (linear, tree-based, and recurrent neural networks) using dispersion data for NO₂, NOx, SO₂, and PM (Particulate Matter), generated by a standard dispersion model (TAPM). The simulated data were obtained for two different scenarios using meteorological data from a mesoscale model and local data from a municipal meteorological station in Santa Margarida i els Monjos, near Barcelona. Gradient Boosting and Random Forest are shown to be the best-performing models. Dispersion and atmospheric stability variables—wind, solar radiation, etc.—prove to be the most significant variables for predicting concentrations. The comparative analysis between the two scenarios addressed revealed remarkably similar performance patterns and variable importance between both cases, although with significant differences in the predictability of PM. Thus, ML-based surrogate models have demonstrated the ability to rapidly and effectively anticipate pollutant dispersion, contributing valuable insights for the development of efficient modelling methods and tools to monitor industrial emissions, anticipate and assess their effects, and support decision-making.