Application Of Machine Learning Technologies In The PandaX Experiments

The Dark matter mystery is one of the most important unsolved problems in the field of particle physics.Despite the abundant evidence supporting the presence of a dark component of mass in the universe,but still,the nature of dark matter remains unknown.However,particle physicists have projposed numerous Dairk Matter candidates such as gravitino,sterile neutrino,neutralino,axion,and WIMP.Among these candidates,WIMP is one of the most appealing candidates for explaining the abundance of Dark Matter.Several methods have been designed for detecting the WIMP such as direct detection,indirect detection,and collider production.Direct detection is the most promising method for detecting WIMP in Earth-based detectors with low backgrounds.PandaX is one of the experiments that is searching for WIMP via direct.detection with liquid xenon dual-phase time projection chamber(TPC),situated at the China Jinping underground laboratory.PandaX experiments are looking for the occasions in which dark matter particles collide with standard model particles and transfer energy to them inside a detector.However,there are several challenges encountered during the experiment including the signal produced in such experiments is very small(1-100keV),which makes it difficult for detecting the signals and obtaining the high precision events.The signals are normally overpowered by tremendous backgrounds,challenging to separate the particular signals from enormous backgrounds.To resolve the aforementioned issues,the employment of new analysis techniques is essential.For instance,PandaX experiments have successfully implemented a boosted decision tree(BDT)method,which is basically a machine-learning technique,to suppress the backgrounds during the PadanaX-Ⅱ analysis.The successful implementation of the BDT technique gives us the motivation to utilize the advanced technology of deep learning in PandaX analysis for suppressing the backgrounds.Recently,deep neural networks(DNNs)and convolution neural networks(CNNs)become the most popular tools in various research fields of particle physics and have the potential to be used in PandaX analysis.PandaX collaboration has made a significant effort to reduce the backgrounds,but different kinds of backgrounds still exist within the detector.The main backgrounds that existed in the PandaX experiment are classified into four categories such as electronic recoil(ER),nuclear recoil(NR),accidental,and surface background.ER is one of the dominant backgrounds due to its high rates relative to the other background sources.Accidental coincidence is one of the challenging backgrounds,which mimics the signature of a real signal(dark matter).NR is also a.dangerous background as the nature and type of such signal are similar to a real signal.The real signal of the WIMPnucleon scattering is NR.Therefore,we will use the single-scattering NR events from the Americium-Bervllium(AmBe)calibration data as a primary input signal source for training deep learning models in the present analysis study.We will use the ER events from the radon calibration data and also the randomly paired events(accidental coincidence)as the primary input background sources for training the different deep learning models in the current analysis study.Mainly two types of physical events are detected inside the detector,one is the NR event and the other is ER event.NR events in the detector are produced by a neutron-emitting source like AmBe.Inside the detector,neutrons interact with xenon nuclei via elastic scattering,which produces an NR event.On the other hand,ER events in the detector are produced by gamma rays.The accidental events,which are not physical events,are produced by random coincidence of isolated S1 and S2 signals.Such type of event occurs in the experiment when the S1 and S2 signals do not come from the same collision events.During the event reconstruction,the unrelated isolated S1 and S2 signals may occur in the same drift window,which leads to the accidental background.The main objective of this thesis is to investigate the feasibility of applying neural network technology in the PandaX experiments.Therefore,we use supervised machine learning techniques to achieve a better distinction between NR events and ER events,and also between NR events and accidental events(non-physical).We conducted a study using deep learning technology for the separation of NR and ER signals using the PandaX-Ⅱ and PandaX-4T data.We proposed a multiple-input convolution neural network model using the NR and ER waveform information of the PandaX-4T experiment to segregate the NR and ER events.We also designed a deep neural network model based on some specific physical features information given to the model as input to separate the NR events from ER.events using the PandaX-Ⅱ and PandaX-4T data.The result demonstrates that we obtained consistent results from these two methods as those of the traditional method.Better separation of NR events from the rest of the background events can provide information about the events that may be possible candidates for dark matter.We demonstrate the use of DNN to suppress the accidental background in PandaX-Ⅱexperiments with a new input,data approach to improve the sensitivity of the PandaXⅡ data analysis.We implemented a DNN model to discriminate between physical and accidental events.We noticed that the experimental result attained with the DNN method,using the old data selection techniqueis improved by 13%in comparison with the BDT method.We also found that the sensitivity result obtained with the new input data method is enhanced by 21.8%relative to the BDT approach.The experimental results showed that the proposed network distinguished the accidental and physical events efficiently compared to existing schemes,and also observed that the sensitivity for dark matter search is improved.This technique can be used in other dark matter experiments to classify accidental and physical events in a more generic way.Moreover,the application of deep learning is not limited to signal and background discrimination in the direct detection experiment.We perform a study using a generative adversarial network(GAN)to generate synthetic datathat matches the same distribution as actual data,which can be utilized in data analysis.We implemented a GAN model using the AmBe dataset of the PandaX-Ⅱ experiment,which generated almost the same fake data samples similar to the real data samples of the PandaX-Ⅱdata.