Update on Neural Network Modeling + T uning Work for FAST Auralee - - PowerPoint PPT Presentation

▶
update on neural network modeling t uning work for fast
SMART_READER_LITE
LIVE PREVIEW

Update on Neural Network Modeling + T uning Work for FAST Auralee - - PowerPoint PPT Presentation

Nov 08, 2022 203 likes •467 views

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018 Update on Neural Network Modeling + T uning Work for FAST Auralee Edelen, Jonathan Edelen, Chip Edstrom IOTA/FAST Department Meeting February 16, 2018 Update on NN studies, IOTA Dept.

slide-1
SLIDE 1

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Update on Neural Network Modeling + T uning Work for FAST

Auralee Edelen, Jonathan Edelen, Chip Edstrom IOTA/FAST Department Meeting February 16, 2018

slide-2
SLIDE 2

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Outline

  • Reminder of Overall Goals
  • Reminder of 2016 Study
  • 2017 Emittance Measurements
  • Update on Simulations
  • Questions and Areas Needing Feedback/Help
slide-3
SLIDE 3

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Overall Motivation

Example images of laser spot (11 Nov. 2017, 10 Aug. 2016) Simulation + Machine NN Model

Optimization

Input

1st Goal: use NN to create a fast-executing, high-fidelity model

from the gun thru CC2

Why do this:

  • Hybrid a priori / empirical model
  • Combine expected and real machine behavior
  • Include input from virtual cathode image direction
  • Prep for downstream experiments (fast execution)
  • Online prediction of beam characteristics (“virtual diagnostic”)
  • Integral part of training/running a NN tuner
  • online opt. + fast switching between conditions
  • natural extension to phase space manipulation (e.g. s2e prediction of
  • ptimal settings for RTFB transform )
  • deploying this tuner is the 2nd main goal
slide-4
SLIDE 4

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

2016 Initial Simulation Study

  • Gather simulation data from PARMELA scans
  • Variety of input distributions
  • NN predictions after the gun and after CC2

(αx , αy) (εnx , εny) (βx , βy) (Np) (E)

Initial Model Inputs and Outputs (after the gun and CC2)

One PARMELA run with 2-D space charge: ~ 20 minutes Neural network model: ~ a millisecond For gun, MAEs 0.4% -1.8% of the parameter ranges For CC2, MAEs 0.9% - 3.1% of the parameter ranges First successful effort to make a fast-executing, high-fidelity simulation surrogate for an accelerator section using NNs

Main Conclusions

For more info:

  • A.L. Edelen, et al. NAPAC16, TUPOA51
  • Poster + slides from prior dept meetings

and other presentations available on request

Next Steps: validate model with data from next run + expand/improve setup in the meantime

slide-5
SLIDE 5

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Where we are now…

Past: Made Improvements to the Overall Setup

Simulation: Switched to OPAL Using 3D space charge routine More realistic initial laser distributions NN Model: Predicting the output sigma matrix Using locally-connected layers Cropped the VCC image More input variables included (e.g. gun gradient, CC1 + CC2 params)

Present: Incorporating Measured Data

Took some emittance measurements after CC2 Checking simulation accuracy + improving Updating the NN model with measured data

Future: Finish Model + Controller Training

Train final model + use to train controller Verify model during next run Verify tuner during next run

slide-6
SLIDE 6

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Conducted Emittance Measurements

  • Multi-slit measurement (w/ Chip’s fitting tool)
  • Scanned over gun phase and solenoid strength
  • Ran for ~130 pc and ~250 pc bunch charge (but this varies over

scans)

  • Things to be aware of for slit measurements:
  • Beam going out of range of slits (size too large)
  • Rotated slit images
  • Double beam à interpretability?
  • Lots of poor fits at higher bunch charge (250 pc)
  • X111 images much cleaner than X120 images for slits at X107
  • Going through data to see for what setting combinations fits are poor
  • First, for fits as-is, see where measurements / simulation match

Lower: X111, 60 pls, GSMI 299, GRESPP 194, ND1

slide-7
SLIDE 7

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Becomes quite messy at x120 with slits at x107

X120, 60 pls, GSMI 299 A right, GSMI 297 A left

slide-8
SLIDE 8

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Comparison of measurements and simulations

  • Measurements of x and y emittance for solenoid and phase scans
  • OPAL simulations at 135 pC and 250 pC for both solenoid scans and phase scans
  • Cavity field maps generated by superfish
  • The solenoid field map is generated by simulating the solenoid and bucking coil with currents from

ACNET

  • The initial transverse distribution was created using the virtual cathode image
  • Comparing measured and simulated projected emittance
slide-9
SLIDE 9

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Auralee Edelen ICHEP 2016

Solenoid model

  • Magnetostatic solution of solenoid and

bucking coil using currents from ACNET and the number of turns from engineering drawings

  • N_GSMI = 300.27
  • N_GSBI = 72.48
  • Note the field on the cathode is

approximately 10% of the peak field

slide-10
SLIDE 10

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Auralee Edelen ICHEP 2016

Phase scan comparison

Normalized X emittance as a function of gun phase for a bunch charge of 135 pC Normalized X emittance as a function of gun phase for a bunch charge of 250 pC

slide-11
SLIDE 11

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Auralee Edelen ICHEP 2016

Main focus now:

  • Improve sims before using for training data + finish comparison
  • look at solenoid scans + account for change in bunch charge over solenoid scan
  • Haven’t compared with / used NIU’s IMPACT
  • T model for these scans yet

(in process of doing so w/NIU)

  • Chip updated the fitting script + made compatible with previously collected image

data (vs. live data from camera) à re-doing fits

  • Collect/list settings for which the emittance fits are poor + why (e.g. double beam,

partially off slits) à diagnosis of effect + help fitting procedure?

  • Request for feedback + any other measured data or simulation results people

already have collected/analyzed

slide-12
SLIDE 12

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Request for Feedback / Help

  • Emittance measurements for higher bunch charge have poor fits
  • Other emittance measurements available? Just need any inputs you have + time stamp (even rough)
  • Any other beam data would also be useful (after gun, after CC2, between CC2 and cryomodule all useful)
  • Simulation matches ok for phase scans where we trust the slit fits, but this is a narrow view
  • Others who have done simulations:
  • how good is the match quantitatively? If better, what’s different between this model + others?
  • all original slit fits + images are in the elog if you want to compare
  • Double-beam effect
  • Any further info would be useful
  • If can incorporate into simulation (esp. w/ simulation of slits), might be able to use more of the measured

emittance data (NIU working on getting IMPACT

  • T model to match; Dan à how involved is sim w/ slits?)
  • FAST repository for simulation files, simulation results, and comparison with measured data?
  • Could combine ASTRA, PARMELA, OPAL, and IMPACT
  • T models and results in one place
slide-13
SLIDE 13

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Backup Slides

slide-14
SLIDE 14

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

For some settings we do still see two sets of images from the slits

X111, 60 pls, gsmi 299, grespp 194, (left nd1, right nd2)

slide-15
SLIDE 15

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

This becomes really quite messy at x120 (297 A and 299 A on the solenoid respectively) X120, 60 pls, gsmi 299 right, gmsi 297 left

slide-16
SLIDE 16

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

When increasing the gain a lot on the camera, we see some ring-like structure in the sparser part of the beam 60 pls, nd2, gsmi 297 left 299 right, grespp 194

slide-17
SLIDE 17

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

For some gun solenoid/phase settings, we see two other portions that look like they've rotated after going through the slits relative to the larger observed group; we took images with Q106 on and off and at x111 and x120 for comparison x120, 60 pls, gsmi 297, grespp 214, UVWP 25%, skew quad on left, off right

slide-18
SLIDE 18

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

x111, 60 pls, gsmi 297, grespp 214, UVWP 25%, skew quad on left, off right

slide-19
SLIDE 19

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

VC image

slide-20
SLIDE 20

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

slide-21
SLIDE 21

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

  • N:GRESPP ~ 170-215
  • N_GSMI ~ 270-310
  • N:T102B
slide-22
SLIDE 22

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

slide-23
SLIDE 23

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

Initial Simulation Data

  • PARMELA simulations from the gun up to the exit of CC2
  • 2-D space charge routine
  • Scanned gun phase, solenoid strength, initial beam distribution
  • Two sets of data:
  • Fine scans (5° phase, 5% sol. str.) for sims just past the gun
  • Coarse scans (10° phase, 10% sol. str.) for sims up through CC2
  • Simulated “virtual cathode images”
  • Going from VCI à initial beam distribution ok from prior work
  • Initial beam distribution à simulated VCI probably ok
  • Obviously very “well-behaved” examples

For normalized sol strength, 1 is the setting that produces a peak axial field of 1.8 kG

Simulation predictions after CC2. Dashed lines are x- emittance, solid lines are y-emittance. Caveat: doesn’t take into account coupling…later changed NN setup to predict sigma matrix, and also used a 3D space charge routine.

slide-24
SLIDE 24

Update on NN studies, IOTA Dept. Meeting, Feb 16, 2018

NN Model: T wo Representative Plots

T

  • p-hat initial beam, 0° RF phase, after gun

Asymmetric Gaussian initial beam, 0° RF phase, after CC2

Dashed lines are NN predictions and solid lines are simulation results

For all gun data, all MAEs are between 0.4% and 1.8% of the parameter ranges. For all CC2 data, all MAEs are between 0.9% and 3.1% of the parameter ranges.

à Surprisingly good for such a small training set