Dynamically Adaptive Numerical Weather Prediction: Models, observations and cyberinfrastructure responding to the atmosphere

After some five decades of operational utilization, numerical weather prediction (NWP) systems (i.e., data ingest, quality control, assimilation, modeling) continue, for the most part, to be run in fixed configurations on fixed schedules. However, increasing computational capability continues to drive such systems toward finer spatial fidelity, giving them the ability to explicitly represent locally intense weather, such as deep convective storms, which are highly localized, intermittent, rapidly evolving, and frequently develop with little forewarning of timing or location. Unfortunately, most technologies used to observe the atmosphere, predict its evolution, and compute, transmit and store information about it operate not in a manner that accommodates the dynamic behavior of such weather, but rather as static, disconnected elements. Radars do not adaptively scan specific regions of storms, numerical models mostly are run on fixed time schedules in fixed configurations, and cyberinfrastructure does not allow meteorological tools to operate on-demand, change their mode in response to weather, or provide the fault tolerance needed for rapid reconfiguration. As a result, today's weather technologies are far from optimal when applied to any particular situation.

In an effort to explore strategies for overcoming some of these limitations, the National Science Foundation funded in 2003 a 5-year, $11.25M project known as Linked Environments for Atmospheric Discovery (LEAD). Involving more than 100 researchers across nine institutions, LEAD has created an integrated, scalable system that allows for the operation of meteorological resources, and associated cyberinfrastructure, as dynamically adaptive, on-demand, grid-enabled systems that can a) change configuration rapidly and automatically in response to weather; b) respond to decision-driven inputs from users; c) initiate other processes automatically; d) steer remote observing technologies, such as Doppler radars, to optimize data collection for the problem at hand; and e) provide the fault tolerance necessary to achieve required levels of performance.

In this paper we present results from idealized numerical simulations of deep convective storms in which we pose the following question: Given fixed computational resources and a given weather scenario, what configuration (e.g., grid spacing, number of nested grids, domain size, start and stop times, observations, physics options, etc) of a numerical prediction model will yield the “best” or most optimal forecast? Our experiments are conducted with the WRF model, and a variety of objective and subjective measures are used to assess forecast quality and utility. By spanning a clearly defined parameter space, we are able to evaluate a cost function that can be used to automatically configure models in a dynamically adaptive framework such as LEAD.