使用宇宙学模拟的笛卡尔切片图作为卷积神经网络的输入
Using Cartesian slice plots of a cosmological simulation as input of a convolutional neural network
AI总结:
本文提出用宇宙学模拟的笛卡尔切片图训练CNN,通过Hessian矩阵特征值标记宇宙网结构类别,验证了二维切片图足以识别宇宙网结构。
摘要:
通过立方体单元的统一划分,我们覆盖了一个基于粒子的ΛCDM宇宙学模拟的总体积。我们将可视化单元定义为立方体单元的空间扩展,从而收集该可视化单元内包含的所有模拟粒子,以创建一系列笛卡尔图,在这些图中物质的过密度清晰可见。然后,我们将这些图作为基于Keras库和TensorFlow的卷积神经网络(CNN)的输入,用于图像分类。为了给每个图分配一个类别,我们在立方体单元中心近似引力势的Hessian矩阵。根据对Hessian矩阵进行Householder约化所获得的正特征值数量,每个选定的立方体单元被赋予标签1、2或3。我们将CNN应用于多个模型,包括两个具有不同可视化体积的模型,一个单元尺寸为L型(大),另一个为S型(小)。第三个模型结合了前述L型和S型单元的图。到目前为止,我们主要考虑平行于XY平面的切片来制作图。最后一个模型基于单元的可视化,其中还包括平行于ZX和ZY平面的切片。我们发现,对图进行分类的准确率是可接受的,并且模型预测类别的能力表现良好。这些结果使我们能够证明本文的目标,即通常的笛卡尔图包含足够的信息来识别所观测到的宇宙网结构。
英文摘要:
Using a uniform partitioning of cubic cells, we cover the total volume of a $Λ$CDM cosmological simulation based on particles. We define a visualisation cell as a spatial extension of the cubic cell, so that we collect all simulation particles contained in this visualisation cell to create a series of Cartesian plots in which the over-density of matter is clearly visible. We then use these plots as input to a convolutional neural network (CNN) based on the Keras library and TensorFlow for image classification. To assign a class to each plot, we approximate the Hessian of the gravitational potential in the centre of the cubic cells. Each selected cubic cell is then assigned a label of 1,2 or 3, depending on the number of positive eigenvalues obtained for the Householder reduction of the Hessian matrix. We apply the CNN to several models, including two models with different visualisation volumes, one with a cell size of type L (large) and the other with a cell type S (small). A third model that combines the plots of the previous L and S cell types. So far, we have mainly considered a slice parallel to the XY plane to make the plots. The last model is considered based on visualisations of cells that also include slices parallel to the ZX and ZY planes. We find that the accuracy in classificating plots is acceptable, and the ability of the models to predict the class works well. These results allow us to demonstrate the aim of this paper, namely that the usual Cartesian plots contain enough information to identify the observed structures of the cosmic web.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。