BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:0
UID:UW-Physics-Event-9622
DTSTART:20260126T180000Z
DTEND:20260126T190000Z
DTSTAMP:20260413T084952Z
LAST-MODIFIED:20260307T173052Z
LOCATION:EH 1227
SUMMARY:Efficiently Learning Linear System Solvers for Fast Numerical 
 Simulation\, Plasma Physics (Physics/ECE/NE 922) Seminar\, Professor M
 isha Khodak \, University of Wisconsin-Madison
DESCRIPTION:Accelerating PDE solving is an important emerging AI appli
 cation\, but popular approaches that fully replace classical solvers u
 sing neural networks often struggle to compete due to insufficient dat
 a\, optimization issues\, low precision\, and a lack of guarantees. We
  consider the alternative paradigm of integrating learning directly in
 to solvers\, focusing specifically on initial value PDEs\, for which t
 he main computational cost is often solving a sequence of linear syste
 ms. We introduce PCGBandit\, a lightweight online learning algorithm t
 hat has performance guarantees under practically reasonable distributi
 onal assumptions on the linear systems' target vectors\, and implement
  it in the popular open-source software OpenFOAM. In evaluations acros
 s six different settings\, including two MHD simulations\, PCGBandit y
 ields significant wallclock reductions while inheriting the classical 
 solvers' precision and correctness. Lastly\, we highlight several futu
 re directions for analyzing scientific computing via the lens of learn
 ing theory/online algorithms and for further data-driven impact on num
 erical simulation.
URL:https://www.physics.wisc.edu/events/?id=9622
END:VEVENT
END:VCALENDAR