1 of 16

Machine Learning for Beam Losses

​

​

Randy Thurman-Keup

Fermilab

​

Alan Turing

2 of 16

Outline

  • No ML Theory (since I am an amateur)
  • Introduction to Loss Monitoring
    • Devices
    • Fermilab Environment
  • Main Injector / Recycler Implementation
  • Results

​

R. Thurman-Keup --- USPAS Hampton, VA

2

February 1, 2024

3 of 16

Concept

R. Thurman-Keup --- USPAS Hampton, VA

3

February 1, 2024

if ( … ) then …

if ( … ) then …

if ( … ) then …

if ( … ) then …

if ( … ) then …

if ( … ) then …

Linear Coding

Known Data

Neural Network

Training Models

4 of 16

Outline

  • No ML Theory (since I am an amateur)
  • Introduction to Loss Monitoring
    • Devices
    • Fermilab Environment
  • Main Injector / Recycler Implementation
  • Results

​

R. Thurman-Keup --- USPAS Hampton, VA

4

February 1, 2024

5 of 16

Beam Loss Monitors (BLM)

  • Wide variety
    • Ionization Chamber (remember Day 2?)
      • No Gain, microsecond signal time
    • Scintillator with Photomultiplier Tube (PMT)
      • Gain, 10’s of nanosecond signal time
    • Neutron Detector
      • Special material needed to convert neutrons to a charged particle or photon
      • Signal times depend on detection device
  • Fermilab uses mostly ionization chambers
    • Generally robust and radiation hard
    • Concentric cylinders

R. Thurman-Keup --- USPAS Hampton, VA

5

February 1, 2024

Glass tube with Ar gas

Outer cylinder

Inner cylinder

6 of 16

Fermilab Main Injector / Recycler Ring

  • The Main Injector and Recycler Rings are both in the same tunnel
  • Only one set of beam loss monitors

R. Thurman-Keup --- USPAS Hampton, VA

6

February 1, 2024

7 of 16

MI/RR Environment

  • Beam can be present at the same time
  • Losses from each machine overlap
  • There are 7 beam loss monitor systems with >260 loss monitors
  • Use Edge AI to separate the losses

R. Thurman-Keup --- USPAS Hampton, VA

7

February 1, 2024

Recycler

Beam

Main Injector

Beam

Beam Loss

8 of 16

Outline

  • No ML Theory (since I am an amateur)
  • Introduction to Loss Monitoring
    • Devices
    • Fermilab Environment
  • Main Injector / Recycler Implementation
  • Results

​

R. Thurman-Keup --- USPAS Hampton, VA

8

February 1, 2024

9 of 16

ML Acquisition System

R. Thurman-Keup --- USPAS Hampton, VA

9

February 1, 2024

10 of 16

ML System Pirate Cards

  • Cards are MitySOM Cyclone 5 FPGAs on custom VME carrier boards
  • Each BLM system has a Pirate card installed in it (7 cards total)
  • All BLM channels are available �to stream simultaneously
  • Data is streamed via UDP in �DDCP protocol format
  • Data frequency is 320 Hz �(current rate of digitizer polling)
  • Each card streams 0.4-0.6 Mb/s �(dependent on number of BLM �channels in crate)

​

R. Thurman-Keup --- USPAS Hampton, VA

10

February 1, 2024

11 of 16

ML System Central Node

  • Intel Aria10 FPGA on a custom carrier board
  • ML model implemented on FPGA
  • UNet ML Architecture Chosen
  • Dual ARM Cores
  • Collects VME reader card streams and aligns
  • Feeds samples to ML model on FPGA
  • Accept inferences from ML model and stream to Redis

​

R. Thurman-Keup --- USPAS Hampton, VA

11

February 1, 2024

12 of 16

Outline

  • No ML Theory (since I am an amateur)
  • Introduction to Loss Monitoring
    • Devices
    • Fermilab Environment
  • Main Injector / Recycler Implementation
  • Results

​

R. Thurman-Keup --- USPAS Hampton, VA

12

February 1, 2024

13 of 16

Results 1

R. Thurman-Keup --- USPAS Hampton, VA

13

February 1, 2024

14 of 16

Results 2

R. Thurman-Keup --- USPAS Hampton, VA

14

February 1, 2024

15 of 16

Conclusion

R. Thurman-Keup --- USPAS Hampton, VA

15

February 1, 2024

“Experimental 3-dimensional tracking of the dynamics of a

single electron in the Fermilab Integrable Optics Test

Accelerator (IOTA)”, A. Romanov et al.

Synchrotron Radiation Images of “profile” of single electron

16 of 16

Extras

R. Thurman-Keup --- USPAS Hampton, VA

16

February 1, 2024