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:20260522T162633Z
LOCATION:Bldg. 6 - Room 102
DTSTART;TZID=Europe/Stockholm:20260701T120000
DTEND;TZID=Europe/Stockholm:20260701T123000
UID:submissions.pasc-conference.org_PASC26_sess175_pap128@linklings.com
SUMMARY:Statistical Equivalence of AI Emulators and Earth System Models: A
  Large Ensemble Study with Ultra-Low-Resolution E3SM
DESCRIPTION:Salil Mahajan, Michael Kelleher, and Ming Fan (Oak Ridge Natio
 nal Laboratory)\n\nWe evaluate the statistical fidelity of a very large en
 semble of an AI/ML emulator, FourCastNetv1, by evaluating it against a sim
 ilarly large ensemble of an ultra–low–resolution configuration of E3SMv3 f
 or forecasts up to 10-day lead time. FourCastNetv1 is trained on this E3SM
 v3 configuration, and initial conditions for FourCastNetv1 forecasts are t
 aken directly from trajectories of the E3SMv3 ensemble. We compare the emu
 lator-generated and E3SMv3 ensembles for RMSE growth, multivariate error c
 ovariance, and extremes across the 20 prognostic variables emulated by Fou
 rCastNetv1. FourCastNetv1 ensembles are found to be strongly underdispersi
 ve: ensemble spread grows robustly with lead time in E3SMv3 while remainin
 g much smaller in FourCastNetv1. For small initial perturbations (e.g. $O(
 10^{-4}K$) in the temperature field), FourCastNetv1 produces essentially n
 o spread and no dispersion growth, in strong contrast to E3SMv3. A nonpara
 metric permutation test on the 20-dimensional RMSE covariance matrices sho
 ws that the covariance and correlation structures of the emulator and E3SM
 v3 differ significantly, with the observed Frobenius-norm statistic lying 
 far in the tail of the permutation null distribution. Across variables, Fo
 urCastNetv1 also under-populates the upper tails of the ensemble distribut
 ions, indicating an under-representation of rare but dynamically important
  events. Our results indicate that AI/ML emulators of global atmosphere mo
 dels, while offering orders-of-magnitude speedups for ensemble prediction,
  require explicit evaluation of ensemble spread, dependence structure, and
  extremes to establish their suitability as replacements for their parent 
 dynamical systems.\n\nSession Chair: Thorsten Kurth (NVIDIA Inc.)\n\n
END:VEVENT
END:VCALENDAR