Author: Frey, M.
Paper Title Page
SUPAF11
Computer Architecture Independent Adaptive Geometric Multigrid Solver for AMR-PIC  
 
  • M. Frey, A. Adelmann
    PSI, Villigen PSI, Switzerland
 
  Funding: SNSF project 200021159936
The ac­cu­rate and ef­fi­cient sim­u­la­tion of neigh­bor­ing bunch ef­fects in high in­ten­sity cy­clotrons re­quires to solve large-scale N-body prob­lems of O(109…10zEhNZeHn) par­ti­cles cou­pled with Maxwell’s equa­tions. In order to cap­ture the ef­fects of halo cre­ation and evo­lu­tion of such sim­u­la­tions with stan­dard par­ti­cle-in-cell mod­els an ex­tremely fine mesh with O(108…109) grid points is nec­es­sary to meet the con­di­tion of high res­o­lu­tion. This re­quire­ment rep­re­sents a waste of mem­ory in re­gions of void, there­fore, the usage of block-struc­tured adap­tive mesh re­fine­ment al­go­rithms is more suit­able. The N-body prob­lem is then solved on a hi­er­ar­chy of lev­els and grids using geo­met­ric multi­grid al­go­rithms. We show bench­marks of a new im­ple­men­ta­tion of an adap­tive geo­met­ric multi­grid al­go­rithm using 2nd gen­er­a­tion Trili­nos pack­ages that ran on Piz Daint with O(104…105) cores.
 
slides icon Slides SUPAF11 [6.094 MB]  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
SUPAF12
Surrogate Models for Beam Dynamics in Charged Particle Accelerators  
 
  • A.L. Edelen
    SLAC, Menlo Park, California, USA
  • D. Acharya, A. Adelmann, M. Frey
    PSI, Villigen PSI, Switzerland
  • N.R. Neveu
    ANL, Argonne, Illinois, USA
 
  High-fi­delity, PIC-based beam dy­nam­ics sim­u­la­tions are time and re­source in­ten­sive. Con­sider a high di­men­sional search space, that is far too large to probe with such a high res­o­lu­tion sim­u­la­tion model. We demon­strate that a coarse sam­pling of the search space can pro­duce sur­ro­gate mod­els, which are ac­cu­rate and fast to eval­u­ate. In con­struct­ing the sur­ro­gate mod­els, we use ar­ti­fi­cial neural net­works [1] and mul­ti­vari­ate poly­no­mial chaos ex­pan­sion [2]. The per­for­mance of both meth­ods are demon­strated in a com­par­i­son with high-fi­delity sim­u­la­tions, using OPAL, of the Ar­gonne Wake­field Ac­cel­er­a­tor [3]. We claim that such sur­ro­gate mod­els are good can­di­dates for ac­cu­rate on-line mod­el­ing of large, com­plex ac­cel­er­a­tor sys­tems. We also ad­dress how to es­ti­mate the ac­cu­racy of the sur­ro­gate model and how to re­fine the sur­ro­gate model under chang­ing ma­chine con­di­tions. [1] A. L. Ede­len et al., arXiv:1610.06151[physics.​acc- ph] [2] A. Adel­mann, arXiv:1509.08130v6[physics.​acc- ph] [3] N. Neveu et al., 2017 J. Phys.: Conf. Ser. 874 012062  
slides icon Slides SUPAF12 [11.505 MB]  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPAG05
Trimcoil Optimisation Using Multi-Objective Optimisation Techniques and HPC  
 
  • M. Frey, A. Adelmann, J. Snuverink
    PSI, Villigen PSI, Switzerland
 
  Funding: SNSF project 200021159936
Un­cer­tain­ties in the bunch in­jec­tion (i.e. en­ergy, ra­dius, ra­dial mo­men­tum and angle) as well as mag­net in­ac­cu­ra­cies harm the isochronic­ity of the PSI 590 MeV Ring Cy­clotron. An ad­di­tional mag­netic field pro­vided by trim coils is an ef­fec­tive so­lu­tion to re­store this con­di­tion. There­fore, an ac­cu­rate de­scrip­tion of trim coils is es­sen­tial to match the turn pat­tern of the ma­chine in sim­u­la­tions. How­ever, due to the high-di­men­sional search space con­sist­ing of 21 de­sign vari­ables and more than 180 ob­jec­tives the turns can­not be matched in a straight­for­ward man­ner and with­out suf­fi­cient HPC re­sources. In this talk we pre­sent a re­al­is­tic trim coil model for the PSI 590 MeV Ring Cy­clotron im­ple­mented in OPAL that was used to­gether with its built-in multi-ob­jec­tive op­ti­mi­sa­tion al­go­rithm to find the 4 in­jec­tion pa­ra­me­ters and the mag­netic field strengths of 17 trim coils. The op­ti­mi­sa­tions were per­formed on Piz Daint (cur­rently 3rd fastest su­per­com­puter world-wide) with more than 1000 cores per job.
 
slides icon Slides TUPAG05 [6.765 MB]  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)