Binary classification of signal and background triggers of a transition edge sensor using convolutional neural networks

TL;DR

This paper explores using CNNs to classify photon signals in a physics experiment but finds traditional methods perform better due to background noise.

Contribution

The study identifies training confusion from black-body radiation as a key limitation for CNN performance in this context.

Findings

CNNs did not outperform cut-based analysis in detection significance.

Black-body radiation from the fiber is the main background source limiting CNN performance.

Regression-based CNNs and structured training data are recommended for future work.

Abstract

The Any Light Particle Search II (ALPS II) is a light shining through a wall experiment probing the existence of axions and axion-like particles using a 1064 nm laser source. While ALPS II is already taking data using a heterodyne based detection scheme, cryogenic transition edge sensor (TES) based single-photon detectors are planned to expand the detection system for cross-checking the potential signals, for which a sensitivity on the order of 10-24 W is required. In order to reach this goal, we have investigated the use of convolutional neural networks (CNN) as binary classifiers to distinguish the experimentally measured 1064 nm photon triggered (light) pulses from background (dark) pulses. Despite rigorous hyperparameter optimization, the CNN based binary classifier did not outperform our previously optimized cut-based analysis in terms of detection significance. Our findings suggest that training confusion, introduced by near-1064 nm black-body photon triggers in the extrinsics background, is a significant factor limiting the CNNs performance for the associated dataset. The fiber coupled black-body radiation was identified as the limiting background source as concluded in our previous works. Given our results, we recommend that future studies explore regression-based CNNs, placing greater emphasis on the use of standardized and carefully structured training data rather than on extensive hyperparameter optimization. While the presented results and associated conclusions are obtained for a TES designed to be used in the ALPS II experiment, they should hold equivalently well for any device whose output signal can be considered as a univariate time trace.

Figures