Automated model selection for the spectral fitting of large samples of active galactic nucleus spectra

Aims . We developed an algorithm to automatically recommend and selected the best model to fit active galactic nucleus (AGN) spectra in the ultraviolet/optical wavelength range, enhancing the efficiency of fitting large samples of AGN spectra by replacing the visual inspection and manual selection of the best model. Methods . We employed the Penalized PiXel-Fitting ( pPXF ) software to fit AGN spectra using a complete model that includes: narrow and broad emission lines (NELs and BELs), Balmer continuum, Balmer high-order emission lines (H8-H50), FeII pseudo-continuum, AGN continuum, and stellar populations for objects with z < 1; we call this model-1. The fit residuals were analyzed using the discrete wavelet transform (DWT), looking for deviations in the DWT coefficients above some empirically determined threshold value. When deviations were detected in regions of interest of a spectrum (i.e., around H α , H β , Mg II , C IV , and [O III ]λ4959, 5007), the significance of the residuals and kinematics were used to recommend (or not) the addition of an extra fitting component for the model. When a new model is recommended, we compared the new fit and the previous one using the root-mean-square (RMS) difference and F-test. The final results of the developed algorithm are the selection of the best-fit model and the corresponding fit results. We validated the results of the algorithm using a sample of 800 SDSS AGN spectra. Each object was fitted using model-1 and the results were visually inspected by three human validators. The validators also recommended (or not) the addition of the same additional components that the algorithm is equipped to recommend. Results . Comparing the recommendation of each validator and the algorithm, the median coincidence fraction is 0.83-0.88 for different threshold values. These values are comparable to the median coincidence fraction of 0.90 between human validators. The computational time of the model recommendation routine was