Machine learning methods for generation of random images and equilibrated configurations of gluon fields in Quantum Chromodynamics

量子色动力学中随机图像生成和胶子场平衡配置的机器学习方法

• 批准号: 2889923
• 金额: --
• 依托单位: University of Liverpool
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2023
• 资助国家: 英国
• 起止时间: 2023 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2889923
• 关键词: Machine; learning; methods; generation; random

Simulations of Quantum Chromodynamics are among the most challenging computational problems in physics, and typically use parallel computing, hardware acceleration, advanced statistical analysis of very large datasets, and low-level code optimization. One of the most important challenges for Monte-Carlo simulations is to equilibrate and de-correlate configurations of gluon fields, so that their stationary distribution reproduces the Boltzmann weight of QCD. This requires complex, non-local updates of gluon fields.I utilise neural networks such as Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) to learn the most efficient non-local updates that lead to fastest equilibration and thermalization of ensembles of quantum fields. I evaluate the use of these methods vs traditional methods in terms of computing resources.I then apply these algorithmic developments to address outstanding questions about possible exotic phases of hadronic matter, such as possible superfluidity and colour superconductivity of extremely dense nuclear matter in neutron star cores.I study the dependence of transport coefficients such as electric conductivity and viscosity on fermion density within two-colour QCD at nonzero quark density. Transport coefficients directly feed into hydrodynamic simulations of quark-gluon plasma, and play an important role in the interpretation of experimental results on hadron and lepton production in heavy ion collision experiments on LHC and RHIC colliders. I apply Monte-Carlo methods used in lattice QCD to study condensed matter systems with strongly correlated electrons

量子色动力学的模拟是物理学中最具挑战性的计算问题之一，通常使用并行计算、硬件加速、对非常大的数据集进行高级统计分析和低级代码优化。对于蒙特卡罗模拟来说，最重要的挑战之一是平衡和解除胶子场的构型，使它们的平稳分布再现量子cd的玻尔兹曼权值。这需要对胶子域进行复杂的非局部更新。我利用神经网络，如变分自编码器（VAEs）和生成对抗网络（GANs）来学习最有效的非局部更新，从而导致量子场集成的最快平衡和热化。我从计算资源的角度来评估这些方法与传统方法的使用。然后，我应用这些算法的发展来解决关于强子物质可能的奇异相的突出问题，例如中子星核心中极致密核物质可能的超流动性和彩色超导性。研究了非零夸克密度下双色QCD中电导率和黏度等输运系数与费米子密度的关系。输运系数直接输入到夸克-胶子等离子体的流体动力学模拟中，在解释LHC和RHIC重离子碰撞实验中强子和轻子产生的实验结果中起着重要作用。本文应用晶格QCD中使用的蒙特卡罗方法来研究具有强相关电子的凝聚态体系