Аңдатпа
Large ensembles of atmospheric dispersion simulations are necessary for uncertainty quantification, source attribution, and emergency response, however, physics-based models like HYSPLIT might be too computationally expensive for operational use. In addition, they have not fully benefited from recent advances in artificial-intelligence-based surrogate modelling, which offer the potential to reduce this computational burden. This study introduces the Dual-Head Mass-Consistent Plume Surrogate (DM-CPS), a machine-learning surrogate to promote the prediction of PM2.5 transport at a much higher speed without compromising the geometry and integrated pollutant mass of the plume. The proposed framework utilizes two types of complementary histogram gradient-boosting models: one trained in the logarithmic concentration domain to reconstruct the plume morphology, and one trained in the original concentration domain to estimate the total pollutant burden. The outputs of the two branches are then combined through a field-level mass-consistency scaling process that preserves the spatial structure of the plume while enforcing a physically consistent integrated pollutant mass. The model is trained on a GDAS-driven ensemble of 96 HYSPLIT simulations over Almaty, Kazakhstan, which includes 1.88 million grid-cell records and physics-informed wind-aligned features. DM-CPS achieves the highest operational dispersion skill (FAC2 = 0.57) with a significant reduction in geometric bias (MG: 2.75 → 1.68) and an improvement in the median integrated-mass ratio (0.13 → 0.23) while maintaining plume structure (SSIM ≈ 0.83) compared to random forest, multilayer perceptron, and conventional gradient-boosted baselines. The surrogate predicts an entire concentration field in about 35 ms, a factor up to five times faster than running HYSPLIT. The cross-regime evaluation also shows a significant gap of generalization in the presence of unseen, out-of-training meteorological data, again emphasizing the need for increased diversity of meteorological training data for operational use. The results show that physically informed surrogate modeling can significantly speed up the prediction of atmospheric dispersion, which is fundamental to run large ensembles.