The goal of Source Term Estimation (STE) is to accurately identify the parameters that describe the source of a release, namely the position and strength . This requires a reliable dispersion model. Recent advancements have integrated machine learning with the Gaussian dispersion model, enhancing the prediction of pollutant concentrations while mitigating the impact of complex terrains. However, pollutant dispersion varies significantly across different atmospheric stability classes. Addressing this, we introduce a novel strategy termed the Multiple Learning Model (MLM), which segments predictions based on stability classes: Neutral ( ), Unstable ( ), and Stable ( ). This approach, by building models to specific atmospheric conditions, promises more precise source estimations. In comparative study, MLM not only refined prediction accuracy from 0.04 to 0.06 but also improved source location estimates, narrowing the discrepancy to 7 m–25 m from the actual source, a marked improvement over traditional random forest models. This methodological advancement underscores the potential of stability-class-specific models in enhancing the accuracy and reliability of pollutant source estimations.