- Lo-Ellen Park Secondary School
- Brendon Matusch
- Dark Matter, Neural Networks, Deep Learning
In all experiments for detection of WIMP dark matter, it is essential to develop a classifier that can distinguish potential WIMP events from background radiation. Most often, clas- sifiers are developed manually, via physical modeling and empirical optimization. This is problematic for two reasons: it takes a great deal of time and effort away from developing the experiment, and the resulting classifiers often perform suboptimally (which means that a greater amount of expensive run time is required to obtain a confident experimental result).
Machine learning has the potential to automate this and accelerate experimentation, and also to detect patterns that humans cannot. However, two major challenges, which are shared among several dark matter experiments, stand in the way: impure calibration data, which hinders training of models, and unpredictable physical dynamics within the detector itself.
My objective was to develop a set of machine learning techniques that address these two problems, and thus more efficiently generate highly accurate classifiers.
I was able to obtain raw data for two dark matter experiments which exhibit these challenges: the PICO-60 bubble chamber , and the DEAP-3600 liquid argon scintillator . For each experiment, I developed and compared three general-purpose algorithms intended to resolve its inherent challenge (impurity and unpredictable dynamics, respectively).
In PICO-60, background alpha and WIMP-like neutron calibration datasets are used for training; however, there is an impurity of 10% alphas in the neutron set. While a conventional classifier was developed (and is believed to be 100% accurate), machine learning in the form of a supervised neural network (NN) has also been previously explored, because of the benefits of automation. Unfortunately, it achieved a mean accuracy of only 80.2% – not usable as a practical replacement for conventional methods in future iterations of the experiment.
In DEAP-3600, photons are absorbed by a wavelength shifting medium and re-emitted in an unpredictable direction, before being detected by one of 255 photomultiplier tubes (PMTs) around the spherical detector. The randomness severely limits the accuracy of conventional classifiers; in a simulation, the best so far removes 99.6% of alpha background, while also (undesirably) removing 91.0% of WIMP events. Because of physical limitations, simulated data is used for calibration, with 30 real-world experimental events available for testing.
I have written a research paper  about my work on PICO-60, which has been approved by the PICO collaboration and pre-published at https://arxiv.org/abs/1811.11308. It is currently undergoing peer review for publication in Computer Physics Communications.
All PICO researchers are listed on my paper for their work on the original PICO-60 experi- ment. They did not contribute to this study; I completed and documented it independently.