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:20260421T090513Z
LOCATION:Bldg. 6 - Room 003
DTSTART;TZID=Europe/Stockholm:20260629T150000
DTEND;TZID=Europe/Stockholm:20260629T153000
UID:submissions.pasc-conference.org_PASC26_sess103_msa216@linklings.com
SUMMARY:Scaling and Scarcity: Developing Robust ML Surrogates for Next-Gen
 eration Fusion Workflows
DESCRIPTION:Samy Kerboua-Benlarbi (EPFL); Christoph Angerer (NVIDIA Inc.);
  Florian Cabot and Francesco Carpanese (EPFL); Jonathan Citrin (DeepMind);
  Hammam Elaian (EPFL); Cyrille Favreau (NVIDIA Inc.); Federico Felici (Dee
 pMind); Davide Fransos (NVIDIA Inc.); Philippe Hamel (DeepMind); Holger Re
 imerdes, Cristian Sommariva, and Elena Tonello (EPFL); Felix Yang (NVIDIA 
 Inc.); and Alessandro Pau and Olivier Sauter (EPFL)\n\nSurrogate modelling
  is emerging as a key paradigm in fusion research. It provides fast, relia
 ble software components to accelerate high-fidelity plasma simulators and 
 digital twins. However, their development is intrinsically linked to data 
 availability and parameterization. This work explores data-driven surrogat
 e models for two distinct applications: the CHEASE equilibrium solver and 
 the SOLPS-ITER edge plasma physics code.\n\nFor the data-rich CHEASE envir
 onment, we detail a parallelized data-generation pipeline. Models trained 
 on these large datasets predict spline coefficients rather than full 1D pr
 ofiles. This formulation allows a separate reconstruction tool to enforce 
 smoothness and boundary conditions, improving training convergence through
  a simplified loss function. Conversely, to overcome the limited sample av
 ailability caused by the high computational cost of SOLPS-ITER, extended d
 atabases are built using data-efficient learning strategies. Applying Gaus
 sian Process Regression provides crucial uncertainty quantification. This 
 approach identifies where new simulations are needed to enrich sparse data
 sets before training robust edge plasma models.\n\nBy tailoring machine le
 arning strategies to available data, the proposed methods deliver valuable
  tools to improve the computational efficiency of next-generation plasma s
 imulators. We conclude on the ongoing evolution of both frameworks, specif
 ically transport code coupling and advanced surrogate architectures, highl
 ighting remaining challenges in co-designing fusion digital twins.\n\nDoma
 in: Physics, Computational Methods and Applied Mathematics\n\nSession Chai
 rs: Stephan Brunner (EPFL), Eric Sonnendruecker (Max Planck Institute for 
 Plasma Physics), and Florian Hindenlang (Max Planck Institute for Plasma P
 hysics)\n\n
END:VEVENT
END:VCALENDAR