Simple ML Application https doi org10 1016j nima

Simple ML Application https: //doi. org/10. 1016/j. nima. 2019. 03. 001 1

2

Testing tubes for “GRINCH” Gas RINg Imaging CHerenkov - Part of Hall A Super Bigbite Spectrometer 500+ PMT’s (ET 9125 FLB 17) – very low light conditions (few P. E. ) JMU characterized and selected tubes from 1000+ tubes from BABAR. Filter set so 90% events are pedestals Determine Gain, dark current, peak to valley ratio. Use various PMT models to extract tube characteristics

PMT Models for low light response SER 0(x) is ideal single photon response. Dossi et al. model: Multi photoelectron response Bellamy et al. model: Degtiarenko (Pavel @JLab) et al. model: Model’s complex – often need hand holding for fits to converge 4

Model Comparison 5

Model Comparison Bellamy vs. Dossi 6

ML Motivation Convergence of fit of PMT spectra to model dependent on starting values. Several step process of fixing some parameters for initial fit helps. Automating this works, but 20% of spectra still need hand holding. Can ML help? Input to ANN are the raw ADC spectra (2^12 bins). Train to the average gain of the three models. Authors felt 4096 inputs was too much. Characterize spectrum by finding 20 bin numbers (X_n) that divide the spectrum into equal numbers of counts. Also use histogram values at these X_n as inputs (Y_n). 7

8

Multilayer Perceptron NN 5% 10% 95% 100% 9

private communication – not in paper 10

ANN results Comparison for tubes not used in training. Similar to model agreement Used ROOT’s MLP, No special hardware needed Author ”For my $$$ this was not ground-breaking in terms of the ML innovation but rather a very practical (if small) application thereof. ” 11