J13C.6 Visualizing Data-Driven AI Models to Engage Operational Forecasters

Thursday, 1 February 2024: 9:45 AM
327 (The Baltimore Convention Center)
Jacob T. Radford, CIRA, Fort Collins, CO; and I. Ebert-Uphoff, J. Q. Stewart, R. T. DeMaria, T. Wilson, J. L. Demuth, M. S. Wandishin, J. Duda, A. McGovern, C. D. Wirz, and M. G. Cains

AI-based weather models have the potential to transform the weather forecasting enterprise with their unprecedented speed and competitive skill metrics. However, these models are primarily being developed through private sector efforts, such as Nvidia (FourCastNet), Huawei (Pangu-Weather), and Google (GraphCast), with limited involvement from their target user base - operational weather forecasters. Here, we detail early efforts to facilitate providing AI-based models to weather forecasters, including developing visualization platforms for exploring output and case studies demonstrating where in the forecast process AI-based models could add value given known limitations of AI models.

The first visualization platform we present is public-facing and provides near real-time output for the most prominent open-source AI-based models initialized with GFS initial conditions, which allows for comparison with physics-based models. This website has been used internally by CIRA for several months to subjectively assess the performance of FourCastNet. The second platform is the Dynamic Ensemble-based Scenarios for Impact Decision Support Services (DESI), a tool that is already in use by NOAA and the National Weather Service, with which we use GEFS initial conditions to create small ensembles for each AI-based model.

Using these tools, we present case studies of high-impact synoptic events, emphasizing how AI-based models could be used to supplement physics-based models. For example, the speed of AI-based models can facilitate efficient sensitivity analyses and probabilistic forecasts that better represent the tails of distributions. On the other hand, diagnosing errors is difficult in AI models, as is depicting realistic-looking meteorological features at long forecast lead-times and for mesoscale or precipitation forecasting. While we provide ideas for how to interpret AI-based model output from developer perspectives, it is of the utmost importance that we begin to engage the operational forecast community to better understand how AI-based models are viewed in the context of existing numerical weather prediction models and forecast processes. We will briefly discuss our planned research toward this goal, including the central role that these visualization tools will play.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner