A residual neural network is proposed to predict the sound absorption of an infinite rigidly-backed porous material from a classical two-microphone measurement above a finite porous sample. The network is trained using the microphones' transfer functions generated by a boundary element model (BEM), with a Delany-Bazley-Miki material model as a boundary condition. The network is validated numerically with BEM simulations and experimentally using two-microphone measurements of a baffled porous absorber of dimensions 60 cm×60 cm and 30 cm×60 cm, subject to various source locations. The results indicate that the network can significantly enhance the predictive capabilities of the classical two-microphone method. The suggested approach shows potential for accurately estimating the sound absorption coefficient of acoustic materials in realistic operational conditions.