Abstract:
In April 2019, the Event Horizon Telescope (EHT) Collaboration released the first image of a black hole (BH) shadow. Theoretical models that aim to describe the environments of BHs are complex and highly-dimensional numerical simulations are often needed to outline the problem. While previous work has employed the use of machine learning (ML) algorithms to predict BH shadow model parameters from image data, in this thesis, we assess the suitability of a particular class of ML algorithms, namely self-organising maps (SOM), as a tool to classify simulated BH shadow images. We employ the SOM network PINK, which spatially compares visual input using a flip and rotation invariant similarity measure, to generate a set of representative BH shadow prototypes for a library of simulated images. Using this and the clustered input data parameter distributions, we find that the shadow ring size, which is related to BH mass in the model, is the dominant class determining factor of the images. Other model parameters, especially those that influence the orientation of the shadow on the image plane, were less influential on the clustering given PINK’s flip/rotation invariance. Despite this, PINK may be useful in determining persistent image-plane features of BH shadows for other model parameters, given a constant BH mass, to curate a subset of meaningfully different models that can then be used in more advanced analyses reducing the volume of data needing further consideration.
