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:20260724T151406Z
LOCATION:Bldg. 6 - 103
DTSTART;TZID=Europe/Stockholm:20260701T143000
DTEND;TZID=Europe/Stockholm:20260701T150000
UID:submissions.pasc-conference.org_PASC26_sess162_msa183@linklings.com
SUMMARY:Earth System Foundation Model - Heterogeneous Data Integration and
  Forecasting
DESCRIPTION:Firat Ozdemir (Swiss Data Science Center); Yun Cheng and Salma
 n Mohebi (Swiss data science center); Fanny Lehmann, Simon Adamov, Langwen
  Huang, Leonardo Trentini, Oliver Fuhrer, Torsten Hoelfer, and Siddhartha 
 Mishra (ETH Zurich); Sebastian Schemm (University of Cambridge); Benedikt 
 Soja (ETH Zurich); and Mathieu Salzmann (Swiss data science center)\n\nThe
  talk introduces the Earth System Foundation Model (ESFM), a fully open fo
 undation model tailored for weather and climate modeling tasks. Built upon
  the Swin Transformer backbone of the pioneering Aurora model, ESFM introd
 uces several key extensions that allow it to process diverse and heterogen
 eous datasets under a single unified framework. Its variable encoding sche
 me and training protocols allow the model to natively handle missing value
 s across all spatio-temporal dimensions, processing everything from dense 
 gridded data (like ERA5 and CMIP) to sparse gridded satellite data (MODIS)
  and non-gridded station data. Key architectural features include the use 
 of axial attention to capture inter-variable dependencies, and individual 
 variable tokenization which allows for variable shuffling during training 
 and simplifies the development of new downstream tasks. Furthermore, ESFM 
 employs a simple and cost effective means to convert the deterministic mod
 el into a probabilistic one. Overall, ESFM retains the key strengths and l
 ong-term stability of Aurora while significantly broadening its functional
 ity, which is demonstrated through empirical evaluations and case studies.
 \n\nDomain: Climate, Weather, and Earth Sciences\n\nSession Chair: Xavier 
 Lapillonne (MeteoSwiss)\n\n
END:VEVENT
END:VCALENDAR