Wildland fire modeling is an important but extremely challenging computational forecasting problem, in which uncertainties in numerical weather prediction are compounded by uncertainties in modeling the complex dynamics of wildland fire behavior, the processes span vast spatial and temporal scales, and some processes have a stochastic nature that cannot be deterministically modeled. For these reasons, Dynamic Data-Driven Application System (DDDAS) techniques, a new paradigm beyond current data assimilation techniques in which simulations and measurements become a symbiotic feedback control system, have great potential for advancing this area. DDDAS entails the ability to dynamically incorporate additional data into an executing applications, and in reverse, the ability of an application to dynamically steer the measurement process. Our ultimate objective is to build a coupled atmospheric-wildland fire modeling system based on DDDAS techniques that is steered dynamically by data that includes weather data, surface and aircraft-based fire data, spatially-varying fuel data, terrain, and other data that influence the growth of fires.

The wildland fire model is a nonhydrostatic, anelastic numerical weather prediction model that is two-way coupled to a semi-empirical fire behavior model in which the fire spread rate depend upon factors in the fire environment such as terrain slope, fuel characteristics, and local atmospheric wind. As the fire spreads, it burns fuel, and sensible and latent heat fluxes and smoke are released into the lowest layers of the atmosphere, creating strong and rapidly changing inflow and updrafts that in turn feed back on fire behavior, i.e. the fire can "create its own weather". In this work, we will describe how we have encapsulated the coupled atmosphere-fire model code in a collaborative software framework and been introducing DDDAS concepts into what was a traditional modeling approach. This involves techniques to change the forecast as new data is received, to assimilate data arriving out of order, to modify the standard ensemble Kalman filter, and to use in problems such as this where the error distributions cannot be assumed to be Gaussian.