BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20220812T074334Z
LOCATION:Sydney Room
DTSTART;TZID=Europe/Stockholm:20220627T173000
DTEND;TZID=Europe/Stockholm:20220627T180000
UID:submissions.pasc-conference.org_PASC22_sess127_msa157@linklings.com
SUMMARY:Deep Learning for Weather Prediction and Ensemble Post-Processing
DESCRIPTION:Minisymposium\n\nDeep Learning for Weather Prediction and Ense
 mble Post-Processing\n\nDryden\n\nRecent advances in deep learning have sh
 own great promise in weather and climate tasks. However, the neural networ
 k architectures and training infrastructure are often adapted from "off-th
 e-shelf" approaches for tasks such as image classification. Such approache
 s need to be scaled and tuned for these new tasks, and understanding the i
 mportance of different components is critical. We will present a suite of 
 successes, failures, and the reasons behind them when applying diverse arc
 hitectures, such as U-Nets, attention, and vision transformers, to two imp
 ortant weather tasks: post-processing ensembles from numerical weather pre
 diction systems and medium-term weather prediction. We will also discuss k
 ey aspects for successfully training such models at scale, including optim
 ized I/O ingestion and distributed training. By combining insights from bo
 th areas, we are able to demonstrate improved results on these tasks and p
 rovide guidance for the community.\n\nDomain: Climate, Weather and Earth S
 ciences, Computer Science and Applied Mathematics, Engineering, Physics
END:VEVENT
END:VCALENDAR