Accurately predicting dust-induced visibility reduction remains a major challenge in arid-region forecasting, particularly due to the strong influence of upper-tropospheric dynamics. This study investigates dust events in Western Iran by separating atmospheric conditions into Jet Stream and Non-Jet Stream regimes and applying a hybrid evolutionary framework combining Genetic Algorithms with Support Vector Machines optimized by the Firefly Algorithm (SVM-FFA). Among all tested models, SVM-FFA showed the best performance. Under Jet Stream conditions, the model achieved high predictive accuracy (Training: R-2 = 0.96, RMSE = 46.70 m, MAE = 33.66 m; Testing: R-2 = 0.94, RMSE = 52.82 m, MAE = 39.95 m) with strong efficiency (NSE = 0.93). Uncertainty analysis also indicated relatively stable predictions with a mean uncertainty band of about 245 m and a relative uncertainty of similar to 36% with 94% coverage in the testing phase. In contrast, Non-Jet Stream conditions showed weaker predictability due to stronger local atmospheric variability; however, SVM-FFA remained the best model (Training: R-2 = 0.70, RMSE = 132.11 m; Testing: R-2 = 0.58, RMSE = 319.93 m). The uncertainty analysis for this regime indicated substantially larger variability, with a mean uncertainty band of about 641 m and relative uncertainty approaching 298%, reflecting the stochastic nature of dust processes without Jet Stream-driven synoptic forcing. Explainable AI (SHAP) further confirmed the dominant influence of Jet Stream-related dynamic drivers, particularly vertical motion and pressure-gradient mechanisms, in shaping dust-related visibility reduction. Overall, the results demonstrate that separating atmospheric regimes significantly improves dust-prediction capability and helps bridge atmospheric dynamics with data-driven modeling.