Machine learning is crucial in modern physics for analysing large datasets. One of the goals of the GALAH sky survey is to research the motions and the chemical composition of stars in our galaxy in order to understand its formation and evolution. To this end, it aims to obtain spectra for one million stars, which represents a huge amount of data that cannot be examined manually. Binary stars pose a particular challenge for GALAH, as their presence can lead to misinterpretation of spectra. To avoid this and at the same time to flag these specific spectra as automatically as possible for further study in the future, we have developed several convolutional neural network models for to the classification of stellar spectra into single or binary stars class, using the spectrum classifiers produced in the GALAH project with other methods. Each model was developed with different architectural choices, including different numbers of layers, filter sizes and other hyperparameters. Our experiments showed that layered architectures tend to give better results, especially for binary star identification. We also found that certain stellar parameters, such as temperature and surface gravity, affect the classification accuracy. Additionally, we have shown that spectra with larger differences in radial velocities are more accurately classified. The fraction of correctly classified single stars was higher than the fraction of correctly classified binary stars. Although convolutional neural networks have proven to be a powerful tool for classifying stellar spectra, many questions remain about how exactly the signal indicating a binary star is transformed into the output of the neural network.
|
