Using Cartesian slice plots of a cosmological simulation as input of a convolutional neural network

Abstract

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.