Utilizing Deep Learning And Data Visualization Techniques To Study Detector Event Reconstructions And Related Areas

Many high energy physics experiments,including the Daya Bay experiment collects and process a sea of data which is then used to extract useful physics information,and finally achieve their physics goals through manipulations of these physics information.In many of these experiments,it is desirable that the ratio of the backgrounds to their signals be as small as possible.In the case of the Daya Bay experiment,this would include determining the location and energy of the interaction of the antineutrino with the detector in the form of inverse beta decay(IBD)thereby increasing the efficiency of the event selection cuts.This is an important factor in determining the neutrino oscillation parameter ?13.In this work,we present the usage of machine learning to solve a regression problem in the Daya Bay experiment,in particular using data visualization and deep learning to perform an event vertex reconstruction on Monte Carlo data.By reducing the uncertainty in the vertex and energy reconstruction of an event,the uncertainty in the number of IBD events as collected by the detector can also be reduced correspondingly.Using the deep neural network as developed from deep learning,we have attempted also to reconstruct the energy of the event.We find that the deep neural network can produce a better vertex resolution in some regions of the Daya Bay central detector than the current AdSimple reconstruction method as used by the Daya Bay collaboration.We then show that the resolution of the reconstructed vertex and energy is a submodular function of the information collected from the sensors.In the case of the Daya Bay detector,this information corresponds to the charges as collected by the photomultiplier tube(PMT)acting as sensors.Using deep neural network in combination with a submodular function,we consider the problem of placing the PMT in strategic locations or the identification thereof,such that a detector can be designed to be information-optimized,i.e.that the overlap of useful information between the PMTs are minimized while optimizing the usage of resources.To our knowledge,the usage of deep learning has only begun about two years ago in the field of high energy physics as of the writing of this thesis to tackle classification problems instead of regression problems which is the case here in this thesis.Moreover,the notion of submodularity coupled with the usage of deep learning has not been found in the literatures of the same said field as of date.Recently,we got to know that the MANTISSA project from the Lawrence Berkeley National Lab(BNL)in the USA has reached an agreement with the Daya Bay Collaboration to share the data from the antineutrino detector for use in machine learning.