Study On A Beam Simulation Resampling Method Based On The Generative Adversarial Network For The COMET Experiment

The COMET experiment is an experiment in high energy physics to find new physics beyond the standard model through?-e conversion.The COMET experiment is divided into two phases.COMET Phase-I aims at a single-event sensitivity of 3×10^-15and measurement of the beam background.The design and optimization of COMET experiments need to use many simulation data,but the Monte Carlo full simulation method is time-consuming.To increase the simulation data statistics,it is necessary to resampling the simulation data.Using the traditional histogram sampling method will lose the randomness of sparsely distributed data,so it is proposed to use the Generative Adversarial Network(GAN)in deep learning to resample the data and learn the potential distribution of simulated data.In this article,the Monte Carlo full simulation data in the COMET experiment is input into the network as training data.During the training process,the generator and the discriminator will perform a zero-sum game,and the simulation data will be resampled through the generator.The main difficulties that need to be overcome are:First,the COMET experiment requires high accuracy.At the same time,the GAN algorithm is not so ideal when learning the fine structure of some physical quantities.It needs to use some data preprocessing method for the data distribution characteristics of different physical quantities.Then,it is hoped that a network can generate a variety of particle data,which puts forward specific requirements on the stability of the network and the design of the network itself.The main work is as follows:First,due to the physical and geometric characteristics of the physical data itself,the training data has some special fine structure features.A series of preprocessing operations are carried out on the data for these features,including dimensionality reduction,transformation,and normalization.Secondly,in the network model design,WGAN GP(WGAN)is used to avoid the problems of gradient disappearance and mode collapse during training.In addition,we add conditional information to the network to distinguish different types of data to realize the full simulation of various particle data.Finally,through the qualitative analysis and the quantitative analysis of the time performance of the GAN and the Monte Carlo simulation,it shows that the data obtained by GAN sampling meets the expected requirements in terms of accuracy.It is less time-consuming than the traditional simulation method leads to five orders of magnitude.By comparing the data distribution generated by the three methods of GAN,histogram resampling,and Monte Carlo full simulation,it is shown that GAN is more random in the generation of sparse data than the histogram resampling method.At present,the use of GAN in high-intensity frontier physical experiments is still in the exploratory stage.The results of this article can provide a research foundation for further exploring the application of GAN in particle physics experiments.