Thursday, 1 February 2024: 2:30 PM
327 (The Baltimore Convention Center)
In an operational setting, one of our goals is to use AI tools to help forecasters more efficiently output operational forecast products. At The Weather Company, one of the most time consuming tasks done by some of our forecasters is creating global frontal analyses. Current global fronts are analyzed every three hours, and global frontal forecasts are analyzed daily going out to 120 hours. The team responsible for these analyses spends over 50 hours a week drawing fronts. Reducing the amount of time our forecasters are spending on analyzing frontal positions would allow them to perform higher-value meteorological decision support tasks. By using an AI fronts model developed by Andrew Justin of the University of Oklahoma, TWC forecasters have been able to visualize AI generated frontal regions in real-time while simultaneously analyzing the globe. During his 2023 internship at TWC/IBM, Andrew improved the speed and accuracy of his deep learning AI frontal analysis model and was able to extend it to the global domain. The model was run in real-time on a testbed supported by the NSF AI Institute (AI2ES) and then provided to the TWC teams via our internal WxWall application, allowing for continuous evaluation, feedback, and targeted development. The model was trained over North America and runs using the 25km GFS every six hours. Utilizing the model’s frontal probability regions reduces the forecaster’s time spent deciding on frontal positions in common synoptic scenarios, and helps to create more confidence when analyzing complex synoptic patterns. The model will continue to be trained over the global domain and eventually incorporate other numerical weather prediction models, potentially including TWC’s hourly updating IBM GRAF model. With enough training, it may be possible to use the frontal probability regions to automatically plot fronts and only require forecaster augmentation to deliver high quality analyses efficiently. This would be consistent with the Human Over the Loop forecasting paradigm used successfully to generate other forecast products.

