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:20260522T162632Z
LOCATION:Bldg. 6 - Room 003
DTSTART;TZID=Europe/Stockholm:20260629T120000
DTEND;TZID=Europe/Stockholm:20260629T123000
UID:submissions.pasc-conference.org_PASC26_sess167_pap116@linklings.com
SUMMARY:The Memory Scaling of Reverse-Mode Differentiation in Particle Acc
 elerator Simulations with Space Charge
DESCRIPTION:Arjun Dhamrait, Edoardo Zoni, Axel Huebl, Ji Qiang, and Chad M
 itchell (Lawrence Berkeley National Laboratory); Ryan Roussel (SLAC); Jan 
 Kaiser (DESY); Chenran Xu (Argonne National Laboratory); and Jean-Luc Vay 
 and Remi Lehe (Lawrence Berkeley National Laboratory)\n\nThe recent develo
 pment of \emph{differentiable} simulation codes for particle accelerators 
 has enabled gradient-based workflows that promise finer control and more r
 ealistic modeling of accelerator facilities. However, when using reverse-m
 ode automatic differentiation, the memory usage continuously increases dur
 ing the simulation, and can potentially exceed the available hardware memo
 ry -- especially when costly space charge computation is included. To stud
 y the memory requirements for differentiable simulations, we have implemen
 ted space charge in Cheetah, a PyTorch-based beam tracking code that suppo
 rts reverse-mode differentiation. We find that the memory usage for revers
 e-mode differentiation grows linearly with the number of macroparticles an
 d cells, and that it is proportional to the number of space charge kicks i
 nvolved in the simulation. This general scaling can be used to evaluate wh
 ether a given differentiable simulation is feasible given hardware memory 
 constraints.\n\nSession Chair: Guillaume Houzeaux (Barcelona Supercomputin
 g Center)\n\n
END:VEVENT
END:VCALENDAR
