Ensemble models have been used in meteorological and pollutant transport predictions to improve forecast skill and reduce overall uncertainty due to the inability of observational networks to provide precise initial conditions and due to limitations in model physics. The most prevalent ensemble methods focus on varying the initial conditions, or blending results from multiple models. Ensembles have been shown to extend the time period over which model predictive skill is demonstrated and are capable of quantifying statistical measures of confidence in model results. For short-term forecasts (e.g. 1-2 days), however, errors arising from model formulation and scale limitations tend to be more important. Evolutionary programming (EP), as an alternative, exploits the uncertainties in model formulation and enables the model prediction to be optimized according to which parameters are found to produce the greatest improvement in performance by minimizing model error. By perturbing the inherent model parameters through EP, rather than the initial dependent observed variables, the range of possible solutions can be determined more economically than by initial condition perturbations; and, the model physical limitations leading to short term uncertainty are able to be tuned by the model itself. Savannah River National Laboratory (SRNL) has developed an operational EP mesoscale forecast system by which each model run is used to generate a series of offspring ensembles which are then used to continuously adjust the model formulation towards values that produce the best forecast. Further, incorporation of ensembles of plume models can also identify uncertainty in predictions of pollutant transport and provide improved determination of the location and characteristics of contaminant emissions. Development the SRNL-EP system will be discussed and results compared to existing ensembles with real-world practical applications of this technique.