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:20260421T090515Z
LOCATION:Bldg. 6 - Room 103
DTSTART;TZID=Europe/Stockholm:20260701T140000
DTEND;TZID=Europe/Stockholm:20260701T143000
UID:submissions.pasc-conference.org_PASC26_sess162_msa156@linklings.com
SUMMARY:WeatherGenerator: A Foundation Model for Weather and Climate
DESCRIPTION:Julian Kuehnert (ECMWF)\n\nThe WeatherGenerator project aims t
 o develop a foundation model for the European community that combines a wi
 de variety of datasets to improve a broad range of applications across spa
 tial and temporal scales. The WeatherGenerator model is a highly multimoda
 l transformer capable of ingesting and predicting a diverse range of struc
 tured and unstructured data, including gridded reanalyses (e.g. ERA5), in-
 situ observations (e.g. SYNOP), and data from polar-orbiting and geostatio
 nary satellites. Its encoder maps all physical input fields onto a fixed-d
 imensional, spatially structured latent representation. To make the model 
 scalable, encoding is initially performed locally, which is easily paralle
 lised, before information is aggregated globally. Besides computational ef
 ficiency, this ensures capturing of physically important dynamics in the E
 arth system across different scales. The WeatherGenerator’s core model is 
 trained using a self-supervised student-teacher approach (JEPA/DINO) to fu
 se the various input datasets and create a shared latent representation of
  the physical state. A variety of relevant quantities can be decoded from 
 the latent state using application-specific decoders. Forecasting is imple
 mented by time-stepping in efficiently in the shared latent space. This pr
 esentation will detail the architecture, training methodology, and perform
 ance of WeatherGenerator, demonstrating its capability as a versatile and 
 extensible tool for data-driven Earth system science.\n\nDomain: Climate, 
 Weather, and Earth Sciences\n\nSession Chairs: Xavier Lapillonne (MeteoSwi
 ss); Ilaria Luise (CERN); and Sebastian Schemm (Cambridge University, UK)\
 n\n
END:VEVENT
END:VCALENDAR