YNAO OpenIR  > 南方基地
基于深度学习的无碰撞引力N体数值模拟加速的可行性研究
其他题名Feasibility Study of Accelerating Collisionless Gravitational N-body Numerical Simulations Based on Deep Learning
赵梓成
学位类型硕士
导师龙潜
2021-07-01
学位授予单位中国科学院大学
学位授予地点北京
学位专业天文技术与方法
关键词引力N体数值模拟 深度神经网络 泊松方程
摘要本文目标是验证深度学习方法在加速无碰撞引力N体数值模拟中的可行性。无碰撞引力N体数值模拟对研究星系、暗物质晕以及宇宙大尺度结构的形成和演化都有重要意义。而无碰撞引力N体数值模拟的传统方法的模拟计算非常耗时,限制了模拟宇宙系统的规模。目前模拟的粒子数规模主要在106-1013之间,而更大量级粒子数规模的模拟任务,以目前的数值模拟算法(所需的计算次数较高)和最新的计算机都难以完成。其中目前常用的无碰撞引力N体数值模拟PM系列方法求解势能(解泊松方程)的步骤所需计算次数较多,严重降低了每个时间步长数值模拟的速度,同时也导致了模拟精度难以提升。本文提出使用深度神经网络代替传统方法更快求解势能,以加速无碰撞引力N体数值模拟中的主要耗时部分。应用最新的深度学习方法,实验探索合适的深度神经网络模型和训练策略,最终找到了辅以残差局部结构的Encoder-Decoder整体网络模型结构以及确定了更适用于求解势能任务的损失函数、训练策略等。验证了深度神经网络求解势能的计算时间复杂度为O(N);随着N增大,深度神经网络方法的计算耗时越小于快速傅里叶变换方法或有限差分法;在同等采样率的情况下,精度优于快速傅里叶变换方法;并且小网格数数据训练的模型可以用于大网格数的数据,具有可扩展性;用三维数据训练本文设计的模型结构,也能得到优秀的泊松方程求解器。故无碰撞引力N体数值模拟中,用深度神经网络可以提升PM系列方法中求解势能的速度,从而有效提升整体的模拟速度,或以更小的计算时间代价提高模拟精度。为将来大规模的无碰撞引力N体数值模拟打下基础。
其他摘要The objective of this paper is to verify the feasibility of using deep learning method to accelerate collisionless gravitational N-body numerical simulation.The numerical simulation of collisionless gravitational N-body is of great significance to the study of the formation and evolution of galaxies, dark matter halos and large-scale structures of the universe.However, the traditional method of the collisionless gravitational N-body numerical simulation is very time consuming, which limits the scale of simulated cosmic system.At present, the particle number scale of simulation is mainly between 106-1013, and the simulation task with larger particle number scale is difficult to be completed by the current numerical simulation algorithm (which requires a large number of calculation times) and the latest computer.The PM series method used in the collisionless gravitational N-body numerical simulation to caculate the potential energy(solving Poisson equation) requires a lot of calculation times, which seriously reduces the numerical simulation speed of each time step, and also leads to the difficulty in improving the simulation accuracy.In this paper, a deep neural network is proposed instead of the traditional method to solve the potential energy faster in order to accelerate the time-consuming part in the collisionless gravitational N-body numerical simulation.Using the latest deep learning method, the appropriate deep neural network model and training strategy is explored experimentally.Finally, the overall network model structure of Encoder-Decoder supplemented by residual local structure is found and the loss function and training strategy that are more suitable for solving potential energy task are determined.It is verified that the computational complexity of potential energy for deep neural network is O(N).With the increase of N, the computation time of the deep neural network method is less than that of the fast Fourier transform method or the finite difference method.In the case of the same sampling rate, the accuracy is better than the fast Fourier transform method.And the model trained by small mesh data can be applied to large mesh data which is extensible.An excellent 3D Poisson equation solver can also be obtained by training the model structure designed in this paper with 3D data.Therefore, in the collisionless gravitational N-body numerical simulation, the deep neural network can improve the velocity of solving the potential energy in the PM series method, so as to effectively improve the overall simulation speed,or improving the simulation accuracy with less calculation time cost.It lays a foundation for the future large-scale collisionless gravitational N-body numerical simulation.
学科领域计算机科学技术 ; 人工智能 ; 计算机应用
学科门类工学 ; 工学::计算机科学与技术(可授工学、理学学位)
页数61
语种中文
文献类型学位论文
条目标识符http://ir.ynao.ac.cn/handle/114a53/25490
专题南方基地
作者单位中国科学院云南天文台
第一作者单位中国科学院云南天文台
推荐引用方式
GB/T 7714
赵梓成. 基于深度学习的无碰撞引力N体数值模拟加速的可行性研究[D]. 北京. 中国科学院大学,2021.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于深度学习的无碰撞引力N体数值模拟加速(7232KB)学位论文 开放获取CC BY-NC-SA浏览 请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[赵梓成]的文章
百度学术
百度学术中相似的文章
[赵梓成]的文章
必应学术
必应学术中相似的文章
[赵梓成]的文章
相关权益政策
暂无数据
收藏/分享
文件名: 基于深度学习的无碰撞引力N体数值模拟加速的可行性研究.pdf
格式: Adobe PDF
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。