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:20260421T090512Z
LOCATION:Bldg. 6 - Room 103
DTSTART;TZID=Europe/Stockholm:20260701T150000
DTEND;TZID=Europe/Stockholm:20260701T153000
UID:submissions.pasc-conference.org_PASC26_sess162_msa212@linklings.com
SUMMARY:Climate in a Bottle: Towards a foundation model for climate
DESCRIPTION:Karthik Kashinath (NVIDIA Inc.)\n\nAI emulators offer a path t
 o compressing, boosting limited ensembles, and improving the latency of in
 teracting with petabyte-scale climate prediction data. However, prevailing
  auto-regressive paradigms offer limited flexibility, and are challenging 
 to train on climate time horizons due to drifts, instabilities and compone
 nt-coupling challenges. Conditionally generative models offer an appealing
  alternative. In this context we demonstrate a generative diffusion-based 
 framework—Climate in a Bottle (cBottle)—for emulating global km-scale clim
 ate simulations and reanalysis on the equal-area HEALPix grid. cBottle con
 sists of two model stages: a globally-trained coarse-resolution image gene
 rator that generates 100km (50k-pixel) fields given monthly average sea su
 rface temperatures and solar conditioning, followed by a locally-trained 1
 6x super-resolution stage that generates 5km (12.5M-pixel) fields; global 
 super-resolution is made affordable using an overlapping patch-based multi
 -diffusion. Overall, cBottle shows promise as an emulator across a battery
  of climate model diagnostics, including diurnal-to-seasonal scale variabi
 lity, large-scale modes of variability, tropical cyclone statistics, and t
 rends of climate change and weather extremes. Moreover, cBottle is a step 
 towards a foundation model, by bridging multiple data modalities (reanalys
 is and simulation) with corresponding utility beyond emulation to tasks su
 ch as zero-shot bias correction, climate downscaling, and channel in-filli
 ng. The code is available at https://github.com/NVlabs/cBottle.\n\nDomain:
  Climate, Weather, and Earth Sciences\n\nSession Chairs: Xavier Lapillonne
  (MeteoSwiss); Ilaria Luise (CERN); and Sebastian Schemm (Cambridge Univer
 sity, UK)\n\n
END:VEVENT
END:VCALENDAR