Gaussian Process Regression Bump-hunting with the Heavy Photon Search Collaboration

Dark matter offers an explanation to discrepancies between astrophysical predictions and observations; e.g., the flat rotation curves of galaxies. The Heavy Photon Search (HPS) experiment at the Thomas Jefferson National Accelerator Facility (JLAB) attempts to find evidence of the dark photon, which couples to the photon, acting as a bridge between the standard model and dark matter. HPS is a fixed-target experiment with an electron beam impinging on a tungsten target. It is hypothesized that the electrons can radiate a dark photon that decays into an e+e- pair which can be measured in the detector. Gaussian Process Regression (GPR) models backgrounds in HPS invariant-mass distributions. GPR is a flexible, non-parametric Bayesian approach; it treats the unknown function itself as a random object, taking assumptions about the properties of data (such as smoothness or periodicity) rather than strict functional forms. Because a Gaussian process (GP) models a probability distribution over all plausible functions, it has built-in uncertainty quantification which makes it ideal for noisy and complex datasets. GPR relies on a chosen kernel to sample functions and construct a covariance matrix. Injected signal studies are performed in order to test the GP fit's signal sensitivity. In this presentation, we show preliminary upper limit results from applying GPR to the 2015 and 2016 engineering run invariant mass distributions using a piecewise search method.