|Authors : ||Priyanka Jalan [Aryabhatta Research Institute of Observational Sciences, Manora Peak, Nainital 263002, India],
Vikram Khaire [ndian Institute of Space Science and Technology, Thiruvananthapuram, Kerala 695547, India],
Vivek M [indian institute of astrophysics, Karnataka 560034, India]|
|Abstract : ||We used multi-task learning to design, build, and apply a convolutional neural network (CNN) architecture to assess Lyman-alpha (Lya) absorption in the Lya forest of quasar spectra. Our technique determines the associated HI column density NHI, as well as the velocity dispersion (b), without any explicit modelling or application of the expected line profile for Lya from quantum mechanics. We fine-tuned the CNN model with a custom training set of Lya absorption profiles with varying SNR and resolution, similar to that of the KECK telescope (6.6 km/s). The technique recovers excellent estimates of NHI and b when tested on a held-back validation set. The method produces a low number of false positives and negatives, but it is challenged by systems with exceptionally low or high NHI and b. We intend to apply our CNN model to quasar spectra from KODIAQ data releases in order to generate catalogues of new Lya absorbers and their attributes. Deep neural networks were also used to verify our findings. This work validates the application of deep learning techniques to astronomical spectra for both classification and quantitative measurements.