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:20241120T082410Z
LOCATION:HG F 30 Audi Max
DTSTART;TZID=Europe/Stockholm:20240604T095500
DTEND;TZID=Europe/Stockholm:20240604T095600
UID:submissions.pasc-conference.org_PASC24_sess158_pos122@linklings.com
SUMMARY:P41 - Probabilistic Weather Forecasting through Latent Space Pertu
 rbations of Machine Learning Emulators
DESCRIPTION:Poster\n\nSimon Adamov (ETH Zurich, MeteoSwiss); Sebastian Sch
 emm (ETH Zurich); Oliver Fuhrer (MeteoSwiss); and Reto Knutti (ETH Zurich)
 \n\nThe intrinsic variability of the atmospheric system is historically re
 produced by ensembles of forecasts based on numerical weather prediction. 
 However, the computational cost of running such ensembles based on perturb
 ed initial conditions is prohibitive. Recent advances in machine learning 
 (ML)-based emulators for medium range weather forecasting have opened up n
 ew opportunities. While these methods require large amounts of training da
 ta and are computationally expensive during training, the inference/foreca
 st step is computationally cheap. We propose a novel approach of perturbin
 g pre-trained ML emulators. As it has been suggested that initial conditio
 n perturbations only work to a limited extent in ML emulators, we propose 
 to perturb the latent spaces of these emulators directly, by adding noise 
 to weight tensors. One advantage of this approach is that the perturbation
 s can be applied iteratively. Thereby, the resulting probability distribut
 ion of the ensemble members can be adjusted to serve a specific need. Firs
 t results suggest that introducing such perturbations allows the previousl
 y deterministic emulator to create a probabilistic ensemble weather foreca
 st. These forecasts are thoroughly evaluated and compared against measurem
 ents from MeteoSwiss (the Swiss national weather service). The error growt
 h and propagation of the perturbations are subject to careful analysis.\n\
 nSession Chair: Iva Kavcic (Met Office)
END:VEVENT
END:VCALENDAR