FLAME: Fitting Lyα absorption lines using machine learning

¹ Center for Theoretical Physics of the Polish Academy of Sciences, Al. Lotników 32/46, 02-668 Warsaw, Poland
e-mail: priyajalan14@gmail.com
² Indian Institute of Space Science and Technology, Thiruvananthapuram, Kerala 695547, India
³ Physics Department, Broida Hall, University of California Santa Barbara, Santa Barbara, CA 93106-9530, USA
⁴ Indian Institute of Astrophysics, Koramangala, Bengaluru, Karnataka 560034, India
⁵ Max-Planck-Institut für Astronomie, Königstuhl 17, 69117 Heidelberg, Germany

Received: 27 February 2024
Accepted: 6 May 2024

Abstract

We introduce FLAME, a machine-learning algorithm designed to fit Voigt profiles to H I Lyman-alpha (Lyα) absorption lines using deep convolutional neural networks. FLAME integrates two algorithms: the first determines the number of components required to fit Lyα absorption lines, and the second calculates the Doppler parameter b, the H I column density N_HI, and the velocity separation of individual components. For the current version of FLAME, we trained it on low-redshift Lyα forests observed with the far-ultraviolet gratings of the Cosmic Origin Spectrograph (COS) on board the Hubble Space Telescope (HST). Using these data, we trained FLAME on ∼10⁶ simulated Voigt profiles – which we forward-modeled to mimic Lyα absorption lines observed with HST-COS – in order to classify lines as either single or double components and then determine Voigt profile-fitting parameters. FLAME shows impressive accuracy on the simulated data, identifying more than 98% (90%) of single (double) component lines. It determines b values within ≈ ± 8 (15) km s⁻¹ and log N_HI/cm² values within ≈ ± 0.3 (0.8) for 90% of the single (double) component lines. However, when applied to real data, FLAME’s component classification accuracy drops by ∼10%. Nevertheless, there is reasonable agreement between the b and N_HI distributions obtained from traditional Voigt profile-fitting methods and FLAME’s predictions. Our mock HST-COS data analysis, designed to emulate real data parameters, demonstrates that FLAME is able to achieve consistent accuracy comparable to its performance with simulated data. This finding suggests that the drop in FLAME’s accuracy when used on real data primarily arises from the difficulty in replicating the full complexity of real data in the training sample. In any case, FLAME’s performance validates the use of machine learning for Voigt profile fitting, underscoring the significant potential of machine learning for detailed analysis of absorption lines.