用 Python 來玩粒子物理實驗吧：An Example of Using Deep Learning to Correct Energy Measurement in the Belle II Experiment

My talk will go from popular science of particle physics experiment, then the topic changes to how the particle experimentist do numerical analyis using Python packages. Finally, an example of CNN regression for observed energy correction will be discussed to show that the improvement of numerical analysis in science by using Python machine learning packages. *introduction of particle physics: the physics world composed from elementary particles, and particle experimentists use accelerator or cosmic rays to research their properties. * Belle II experiment: one particle physics puzzle is that most of matter in Universe compsed from particles, not anti-particles. However, we believe that at big bang, number of particles should be equal to anti-particles. So, there are something basic differences than charge sign for particle and anti-particle. Belle II experiment prepare to use accelerator to manufacture ~10^{8} B meson and anti-B meson pairs. B meson is a special particle: Nobel Prize winners-- Kobayashi and Masukawa preidcted the life time of B meson and that of anti-B meson are different. After it decayed, Belle detector measure its daughter particles and try to use its daughter information to recombine B meson and measure its life time. (Belle II experiment wiki:https://en.wikipedia.org/wiki/Belle_II_experiment, public official site: http://belle2.jp/) * Data and Numerical Analysis in Belle II experiment: Belle II detector is a onion structure that is consist of sevaral sub-detectors. Each sub-detector has about thousands of pixels to collect the daughter particles information such as track, ring and cluster. So, we have to build up 0.1 billion B meson events which is composed from 10^{4} pixels. Also, the background events from other particles is thousand times of B mesons. The data perduction (including simulation production) can up to ~1,000MB/sec during the coming operation time. We use many stastical method such as maximun likelihood value, chi-square value, etc to pick up the B meson events from numerous background events. Recently, machine learning started to be used in our field to classify signal or correct the measurement value. * Gamma-ray energy measurement problem: The sub-detector of ECL (electronic calorimeter) in Belle II experiment measure gamma-ray energy with under-estimation. So, this problem cause that we can't measurement some B mesons decays to a gamma-ray daughter precisely. ECL is a two-dimension crystal array that measuring gamma-ray energy, and each crystal record some part of energies from gamma-ray. The energies distribution over 2 dimension array can be a hint for the correction of measured energy (summation of energies at all crystal). As we have no idea about what kind of regression fuction can be used, so CNN regression is a suitable choice to solve this problem. This talk will use tensorflow python API to correct simulated data of smaller detector. In the near future, I hope this CNN method can be deployed to real Belle II analysis.