Training GAN to simulate EXO-200 scintillation signal Shaolei Li - - PowerPoint PPT Presentation

Training GAN to simulate EXO-200 scintillation signal

Shaolei Li DANCE-ML workshop 6 Aug 2020

1

Introductions

• EXO-200 used large avalanche photodiodes (APDs) to measure

scintillation light.

• Simulations of APD signals were difficult and time-consuming and so

far there is not a good method to approach that goal.

• A new approach to fast-simulations is generative adversarial networks

(GANs).

• They produce artificial images from random input while guided by real

images.

• The generator we want is constrained during the training such that the

signals show the expected dependency on the energy and positions.

2

WGAN structure

• Generator: get noise, label,

and generate artificial waveform.

• Critic (discriminator): receive

artificial and real waveform, and give Wasserstein distance.

• Constrainers: supervise

training.

• Keras and Tensorflow are

used to train GAN at SLAC GPUs.

3

Training set

4

• Total107210 events.
• Being flatten in space in
• rder to prevent networks

from cheating.

• Charge energy works as

label during the training.

• Weak Th calibration data

are used to train GAN.

Convergence of constrainers

5

Generated waveform vs. Real waveform

6

cm cm cm

keV

Peak amplitude along z axis

7

• Add up waveforms on one side of cylinder, and select maximum point as the

amplitude.

• Generator receives labels of 2615 keV and random positions, while selecting events

around 2615 200 keV.

• The plots show mean value of all events and the corresponding standard deviation

as error-bars.

±

Plane A Plane B

Training results reconstruction

8

• GAN can also do reconstruction work using constrainer parts.
• Comparing constrainer results after 100 epoch using both GAN and real waveforms and EXO-200 recon results.
• The plots shows testing results using validation set with 6000 events.

Compare DNN and EXO recon Compare DNN recon on GAN and real waveform

Deviations of energy around peak

9

• Select events in 2300~3000 keV, and use labels to create artificial waveform.
• Use constrainer to reconstruct energy from both artificial and real waveforms.
• The histograms shows the deviations between them and means are close to 0.

Compare recon results of general events

10

• 100000 events
• utside training

and validation sets.

Conclusion and further studies

• GAN is able to generate waveforms close to the real ones.
• In order do better training, better constrainers will help.
• We may try changing tensor shape in order to be closer

to the real APDs range on the plane.

11