The application of artificial neural networks ( ANNs ) for the estimation of HI gas mass fraction ( M _ { HI } / { M _ { * } } ) is investigated , based on a sample of 13,674 galaxies in the Sloan Digital Sky Survey ( SDSS ) with HI detections or upper limits from the Arecibo Legacy Fast Arecibo L-band Feed Array ( ALFALFA ) .
We show that , for an example set of fixed input parameters ( g - r colour and i -band surface brightness ) , a multidimensional quadratic model yields M _ { HI } / { M _ { * } } scaling relations with a smaller scatter ( 0.22 dex ) than traditional linear fits ( 0.32 dex ) , demonstrating that non-linear methods can lead to an improved performance over traditional approaches .
A more extensive ANN analysis is performed using 15 galaxy parameters that capture variation in stellar mass , internal structure , environment and star formation .
Of the 15 parameters investigated , we find that g - r colour , followed by stellar mass surface density , bulge fraction and specific star formation rate have the best connection with M _ { HI } / { M _ { * } } .
By combining two control parameters , that indicate how well a given galaxy in SDSS is represented by the ALFALFA training set ( PR ) and the scatter in the training procedure ( \sigma _ { fit } ) , we develop a strategy for quantifying which SDSS galaxies our ANN can be adequately applied to , and the associated errors in the M _ { HI } / { M _ { * } } estimation .
In contrast to previous works , our M _ { HI } / { M _ { * } } estimation has no systematic trend with galactic parameters such as M _ { \star } , g - r and SFR .
We present a catalog of M _ { HI } / { M _ { * } } estimates for more than half a million galaxies in the SDSS , of which \sim 150,000 galaxies have a secure selection parameter with average scatter in the M _ { HI } / { M _ { * } } estimation of 0.22 dex .