Accelerating computations for dark matter direct detection experiments via neural networks and GPUs

Abstract

There is indisputable evidence for the existence of dark matter (DM). Examples are the rotation curves of galaxies, the velocity dispersions of galaxy clusters and dark matter density measurements. One of the biggest questions in physics today con cerns the nature of dark matter, and the most promising theory is that dark matter consists of one or more new particle species. To discover the nature of dark matter particles, they need to be inferred from collider experiments or found via indirect or direct detection. Since none of these alternatives has led to conclusive results within the current theoretical frameworks, new approaches should be investigated. In this thesis, sub-GeV dark matter particles are studied through interactions in direct de tection experiments described with an effective field theory (EFT). More specifically, dark matter-induced electronic transition rates in crystal detectors are studied. The rate of electronic transitions is described with EFT scattering amplitudes, which introduce many model-independent coupling strengths. By computing transition rates corresponding to different sets of EFT parameters, direct detection data can be used for inferring properties of dark matter particles without relying on any specific theoretical framework. Since the computation of the electronic transition rates is very expensive, the aim of this thesis is to implement a deep neural network for fast predictions of transition rates. Furthermore, since the neural network re quires a large data set for training, the generation of training data was accelerated using computations on graphics processing units (GPUs). I developed two neural networks, one with the DM mass as input and one with the DM mass and two EFT coupling strengths as inputs, that are about 600 times faster than the original computations and capture the overall behaviour of the transition rates. However, the relative error of the predictions has a standard deviation of about 30% with a mean of around 0%. On the other hand, the GPU computations are about 16 times faster than the original computations and have negligible error while being able to compute transition rates corresponding to all 28 coupling strengths. I conclude that there is great potential for using both neural networks and GPUs for dark matter research, and suggest further improvements.