The load transfer in pile-supported embankments is complicated, and existing analytical models may not cover all scenarios in real practice due to the assumptions made within them. The 3D numerical method is a more effective alternative to traditional methods. However, numerical simulations require significant computing resources and may encounter convergence problems. In this study, an enhanced model was proposed to evaluate the soil arching effect by combining the well-recognized concentric arching method and a designed neural network. An orthogonal array was established to generate 154 sets of input parameters, which were used as inputs for a series of finite element simulations. The designed neural network is used to bridge the discrepancy between the results of the concentric arching method and that of the 3D numerical simulation. Bayesian information criterion was used to determine the optimal model. The proposed model is validated through measurements from in-situ cases. The results reveal that factors such as the friction angle of the embankment fill, embankment height, and pile spacing significantly affect the soil arching effect. While the extra load has little impact on load transfer efficiency when the embankment height is greater than the pile spacing.