Advancing U.S. Operational
Weather Prediction Capabilities in the Next Decade with Exascale HPC,
Machine Learning and Big Data Technologies
Mark Govett
ESRL/GSD
22 May, Noon, in 2155
Abstract:
A revolution in computing, modeling, software and big data is needed to
advance U.S. weather prediction capabilities in the next decade. An
estimated 1000 to 10000 times more computing is needed to advance
prediction models to cloud-resolving, 1-3KM resolution global scales.
However, existing models are not capable of exploiting future HPC
systems with tens to hundreds of millions of processors. Models will
need to be rewritten to use more efficient algorithms, incorporate
parallelism at all levels, minimize inter-processor communications, and
improve I/O efficiency. In addition, Artificial Intelligence (AI) can
be used to replace computeheavy calculations with fast, light-weight
algorithms. AI and the rapidly growing field of Machine Learning (ML)
has the potential to disrupt the way weather prediction and
assimilation models are developed in the future.
In addition, the current prediction system is being overwhelmed with
too much data. New strategies are needed to more effectively handle the
ingest, processing, computation, and distribution of data within the
prediction system. Emerging technologies such as 5G networks, cloud
computing, ML, and edge computing can be used to support the
processing, distribution, and dissemination of data, information, and
insights to diverse end users.
This presentation will offer a critical and visionary assessment of key
technologies and developments needed to advance U.S. operational
weather prediction in the next decade. I will describe challenges in
our prediction system today and highlight exploratory developments at
GSD and other modeling centers to overcome these challenges.