ML for weather forecasting
Tuesday, June 29, 2021 1:15 PM to 1:35 PM · 20 min. (Africa/Abidjan)
Stream#2
Information
Contributors:
Abstract:
Weather prediction is a demanding large scale scientific challenge that exercises the limits of HPC platforms. Strict latency constraints mean that significant atmospheric processes are unresolved and, in addition, over 95% of observational data is discarded, at substantial cost. In this talk, I'll discuss ML methods for direct weather forecasting, which offer very low latency and high resolution for short range forecasts. ML models like MetNet rely on wide and deep networks to capture atmospheric dynamics up to an 8-12 hour time scale. Low computational costs mean that these ML forecasts can be evaluated ``live'' as observational data arrives, improving forecast quality and observation utilization. We'll also discuss how these methods can augment HPC atmospheric models to improve forecasting at all time scales.
Abstract:
Weather prediction is a demanding large scale scientific challenge that exercises the limits of HPC platforms. Strict latency constraints mean that significant atmospheric processes are unresolved and, in addition, over 95% of observational data is discarded, at substantial cost. In this talk, I'll discuss ML methods for direct weather forecasting, which offer very low latency and high resolution for short range forecasts. ML models like MetNet rely on wide and deep networks to capture atmospheric dynamics up to an 8-12 hour time scale. Low computational costs mean that these ML forecasts can be evaluated ``live'' as observational data arrives, improving forecast quality and observation utilization. We'll also discuss how these methods can augment HPC atmospheric models to improve forecasting at all time scales.
