Improving particle classification in WIMP dark matter detection experiments using neural networks

Brendon Matusch, Canada 13-15

In WIMP dark matter experiments, manually developing a classifier to separate WIMP candidates from background radiation is essential, but challenging and time-consuming. I present novel machine learning algorithms that solve this problem via automation, performing significantly better than previous methods in the PICO-60 and DEAP-3600 experiments. I approached challenges with PICO-60 by developing semi-supervised learning algorithms that alleviate data impurity, improving accuracy from 80.7% to 99.2%. I also present new processes that better handle the DEAP-3600 detector topology, reducing the number of false positives from 91.0% to 75.7%.

Show more