AbstractThe teleseismic receiver function (RF) is commonly used to determine major interfaces of the Earth. If the crustalVpis known approximately, the Moho convertedPsphase and crustal multiple reverberations can be used to determine the thickness (H) and averageVp/Vsratio (κ) of the crust. A widely used method for this isH‐κstacking (Zhu & Kanamori, 2000,https://doi.org/10.1029/1999JB900322), which uses grid search and superposition to find the maximum coherent energy of the MohoPsand its reverberated multiples phases. However, this method assumes a homogeneous isotropic crust and a flat Moho. Furthermore, it is affected by the reference crustalVp. Improved methods, such as theH‐κ‐cmethod for anisotropic media and inclined interfaces (J. Li et al., 2019,https://doi.org/10.1029/2018JB016356), help alleviate the problem. In this paper, we propose a new method that uses deep learning to estimateHandκ. Our method is divided into two steps. The first step employs a denoise architecture (the DenoiseNet) to reduce the noise level of the RFs and restore missing back‐azimuthal information. In the second step, our new deep learning network (the HkNet) is used to estimateHandκ. Deep learning has the inherent ability to automatically extract complex features from RFs, which allows us to estimate the parameters in complex media with different crustalVp. Synthetic data tests show that the proposed method achieves better accuracy than theH‐κandH‐κ‐cmethods. Applications to real data show that the proposed method is robust and reliable in a wide range of geological settings.