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:20250822T115806Z
LOCATION:Room 5.0B15 & 16
DTSTART;TZID=Europe/Stockholm:20250616T115000
DTEND;TZID=Europe/Stockholm:20250616T122000
UID:submissions.pasc-conference.org_PASC25_sess139_msa184@linklings.com
SUMMARY:Advancing Probabilistic Weather Forecast through Machine Learning 
 at Scale
DESCRIPTION:Boris Bonev, Thorsten Kurth, and Mauro Bisson (NVIDIA); Ankur 
 Mahesh (Lawrence Berkeley National Laboratory); Kamyar Azizzadeneshli and 
 Karthik Kashinath (NVIDIA); Michael S. Pritchard (NVIDIA; University of Ca
 lifornia, Irvine); William D. Collins (Lawrence Berkeley National Laborato
 ry); and Anima Anandkumar (California Institute of Technology)\n\nWe prese
 nt recent advancements in global weather modeling based on spherical neura
 l operators. This innovative approach demonstrates superior skill and redu
 ced computational costs compared to current state-of-the-art models. Our m
 odel is trained as a probabilistic ensemble, respecting spherical geometry
  and symmetries. This approach, inspired by first principles and its proba
 bilistic formulation ensure plausible dynamics and stable spectra in the m
 odel's output.\nTo scale training, a novel paradigm for model-parallelism,
  inspired by domain-decomposition is developed. This reduces memory and IO
  requirements per GPU, enabling training of larger models by splitting the
 m across multiple GPUs. Leveraging model-parallelism in conjunction with d
 ata-parallelism and ensemble-parallelism enables massive parallelization t
 o train the model on 1024 NVIDIA H100 GPUs.\nThe resulting model's efficie
 ncy is remarkable, capable of generating a full year's rollout in under 13
  minutes on a single GPU, while demonstrating skill improvement over curre
 nt operational models at 0.25 degrees global resolution.\nThese advancemen
 ts represent a significant step forward in ML weather forecasting, offerin
 g improved accuracy and computational efficiency. The model's ability to c
 apture uncertain dynamics of the weather system while maintaining realisti
 c physical properties makes it a powerful tool for generating forecasts, p
 roviding early warnings for extreme events, and informing decision-making 
 across various sectors.\n\nDomain: Climate, Weather, and Earth Sciences, C
 omputational Methods and Applied Mathematics\n\nSession Chairs: Christian 
 Lessig (ECMWF, Otto-von-Guericke-Universitat Magdeburg) and Ilaria Luise (
 CERN)\n\n
END:VEVENT
END:VCALENDAR