Adviser: Dr. George S. Young, Abstract: We can never quantify the atmospheric state precisely: there will always be an uncertainty associated with measured quantities and model output. This work seeks to understand both the uncertainty introduced by measurements and the uncertainty introduced by approximating the model’s nonlinear terms. We seek both to understand these sources of uncertainty and to incorporate that understanding throughout the scientific process for Atmospheric Transport & Dispersion (AT&D;) problems. First we will examine how errors in the input wind fields may translate into AT&D; model solution errors. We focus on street-level concentration plume errors that occur in building-aware AT&D; models for a set of hazardous release scenarios where the release location varies relative to the building locations and city building configurations. Second, we use Source Term Estimation (STE) techniques to examine how estimates of uncertainty in measurements, e.g. the wind direction, can be used to help bound the problem. We use two techniques to examine the STE problem. Using the Genetic Algorithm coupled with an AT&D; model Variational (GA-Var) method we preform sensitivity analyses to achieve five goals: (1) establish adequate thresholds to filter out noise in our concentration data without decimating the signal; (2) use a robust statistical method to quantify the uncertainty in our predictions; (3) determine the best cost function for each of the variables we seek to retrieve; (4) given that real-time wind direction data are difficult to come by, determine if the GA-estimated wind direction is representative of the advecting wind; and (5) determine the robustness of the GA when a limited number of sensors are available. To further examine the STE problem we use the Variational Iterative Refinement STE Algorithm (VIRSA). VIRSA is a combined modeling system that includes the Second-order Closure Integrated PUFF model, a hybrid Lagrangian-Eulerian Plume Model (LEPM), and its formal adjoint. While numerous approaches to the STE problem exist, each with its own strengths and weaknesses, this approach addresses STE in an operational environment where computational resources are limited and a timely solution is critical. VIRSA is an adjoint method, computationally efficient, and fast but like any gradient descent minimization its downfall is that it can fall prey to local minima in the solution space. In the work presented here we incorporate new methods to address issues related to uncertainty and using what we know about that uncertainty to reduce the tendency to find local minima rather than the global minimum. We explore approaches to map the uncertainty in our observations and link it back to the background error covariance matrix utilized by the adjoint minimization.