BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Data-parallel based deep learning for the model order reduction of
  parametrized partial differential equations    - Dr Nirav Vashant Shah\, 
 CUED
DTSTART:20240216T143000Z
DTEND:20240216T150000Z
UID:TALK211792@talks.cam.ac.uk
CONTACT:46601
DESCRIPTION:Parametrized partial differential equations are used to model 
 problems in engineering. Design of engineering systems is governed by phys
 ical parameters such as material properties\, boundary conditions and geom
 etric parameters such as shape of the components. In order to rapidly expl
 ore the variation in quantity of interest with respect to the physical or 
 geometrical parameters\, model order reduction is used as a computationall
 y faster alternative with an “affordable” compromise in accuracy.\n \n
 Deep learning based model order reduction methods have gained traction in 
 recent years. These methods can be non-intrusive in nature and may not req
 uire access to source code used to solve the high-fidelity model. In the c
 ase of offline-online two stage procedure\, deep learning methods are quic
 ker in the online phase. However\, during the offline phase\, they suffer 
 from severe computational costs associated with generation of training dat
 a and training of artificial neural network. On exascale systems\, such ap
 proaches require more careful numerical implementation due to heterogeneou
 s mixed CPU/GPU devices.\n\nIn this talk\, we will focus on problems invol
 ving geometric parameters. Further\, we will introduce data-parallel distr
 ibuted training of the artificial neural network in order to address the i
 ssue of high offline cost. We will also introduce PyTorch-RBniCSx-FEniCSx 
 based open source package\, DLRBniCSx\, for deep learning based model orde
 r reduction.
LOCATION:Oatley 1 Meeting Room\, Department of Engineering
END:VEVENT
END:VCALENDAR