BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:0
UID:UW-Physics-Event-9069
DTSTART:20250310T170000Z
DTEND:20250310T181500Z
DTSTAMP:20260413T151539Z
LAST-MODIFIED:20250121T195656Z
LOCATION:1227 Engineering Hall
SUMMARY:"Structure-exploiting sparse grid approximations for efficient
  uncertainty quantification and surrogate model construction"\, Plasma
  Physics (Physics/ECE/NE 922) Seminar\, Ionut Farcas\, Virginia Tech
DESCRIPTION:Gyrokinetic simulations on parallel supercomputers provide
  the gold standard for theoretically determining turbulent transport i
 n magnetized fusion plasmas. Applications to large and costly future m
 achines\, in particular burning plasma devices\, call for a proper Unc
 ertainty Quantification (UQ) in order to assess the reliability of cer
 tain predictions. However\, since UQ requires an ensemble of simulatio
 ns\, the high computational cost of gyrokinetic simulations prevents s
 traightforward applications of conventional UQ approaches. To overcome
  this\, we propose a structure-exploiting\, data-driven method based o
 n sparse grid approximations to enable UQ in computationally expensive
  simulations. By leveraging the fact that the quantities of interest (
 e.g.\, heat or particle fluxes) often exhibit strong dependence on onl
 y a subset of the uncertain parameters characterized by anisotropic co
 uplings\, our method significantly reduces the number of expensive sim
 ulations required. We demonstrate this in the context of turbulent tra
 nsport at the edge of tokamaks driven by electron temperature gradient
  (ETG) modes. In a nonlinear scenario with eight uncertain inputs\, ou
 r sparse grid approach requires a mere total of 57 high-fidelity simul
 ations. This efficiency extends to the construction of surrogate trans
 port models\, which are crucial for tasks like the design of optimized
  fusion devices. We will show that our structure-exploiting sparse gri
 d approach can be effectively used to construct a surrogate model for 
 the ETG-driven electron heat flux that delivers predictions with an ac
 ceptable level of precision across a wide range of parameter values. F
 inally\, time permitting\, we will discuss how our data-driven approac
 h can be extended to multi-fidelity methods. By incorporating hierarch
 ies of high- and low-fidelity models\, these methods can significantly
  accelerate computations while maintaining accuracy\, making them part
 icularly promising for complex applications like fusion plasma simulat
 ions.
URL:https://www.physics.wisc.edu/events/?id=9069
END:VEVENT
END:VCALENDAR