Machine Learning techniques to discover and understand the population of flare stars in MeerLICHT data

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title
In this era of information overload, machine learning and artificial intelligence have been increasingly popular in various fields, including the field of astronomy. These approaches attempt to extract meaningful information from the data through automated means. In this work, we develop generic machine learning models that classify a given transient object from the observed light curve. We train random forest (sect 4.1.1) and multilayer perceptron neural network (sect 4.1.3) models on simulated LSST PLAsTiCC data and real data from the MeerLICHT survey. We found that the random forest model outperforms the neural network model in both data sets, achieving test accuracy of 66.0% and 98.0% in the PLAsTiCC and MeerLICHT data respectively. On the other hand, the neural network model achieved test accuracy of 65.7% and 86.6 % in the PLAsTiCC and MeerLICHT data respectively. For PLAsTiCC simulated data, we also show that grouping all types of supernovae into one aggregate class and discarding distance information improves the performance of both models to 96.5% and 96.0% for random forest and neural networks respectively. As additional work, we attempt to find sub-classes within the M-type class in MeerLiCHT data using k-means and hierarchical clustering algorithms. We find two distinct sub-classes in this data. Namely variable and non-variable M-type stars.