Abstract: In this paper, the radial basis function (RBF) method is used to optimize the accuracy of Garvey-Kelson local mass relation (GK) and its generalized formula (GKs and GK+J) in the description and prediction of nuclear mass. A large number of numerical experiments show that the RBF can reconstruct the deviation function with high quality and significantly improve the prediction ability of the model if only reference nuclei at a certain distance around the target nucleus are used. Based on the known nuclear mass data of AME2003, AME2012 and AME2020, the root-mean-square deviation of the predicted results of the three local mass relations can be reduced by an average of 17, 16 and 539 keV, respectively. In the case of extrapolation, it also has the optimization effect for local mass relations. At the end of this paper, the mass excess predictions of 19 nuclei measured by the latest experiments after AME2020 are given (⁶⁵Cr, ^68,69Fe, ^74,75Ni, ⁶⁰Ga, ¹⁰³Sr, ⁸⁰Zr, ¹²⁰Rh, ^99,133,134In, ¹⁰³Sn, ^152-154Ce, ^150,153Yb, ²⁵¹No), and the prediction accuracy of 12 nuclei is better than that in AME2020.