Abstract: To meet the computational efficiency requirements for multi-physics simulations in the design and development of advanced reactor cores, a neural network-based acceleration method was investigated for multi-physics coupling calculations in this paper, building upon the full neutron spectrum code for advanced reactor simulation FSAR. The FSAR code employed a deterministic two-step computational scheme based on homogenization theory for reactor core neutronics analysis, comprising two main components: a cross-section generation module and a core calculation module. A set of one-dimensional single-phase governing equations were solved in the thermal-hydraulic module, including coolant momentum and energy and mass conservation equation, and fuel heat conduction equation. The fuel control equations of the reactor core were solved using a parallel multi-channel model. The radial displacement deformation model for fuel pellets and cladding, along with the Ross&Stoute gap heat conduction model, was employed to account for structural deformation effects. These modules were integrated with the neutronics calculations of FSAR code to achieve nuclear-thermal-mechanical coupling analysis of the advanced reactor core. In this study, a data-driven approach was adopted to construct surrogate models. The fully connected deep neural network was utilized to develop the surrogate model for nuclear-thermal-mechanical coupling calculations, which consisted of an input layer, three hidden layers, and an output layer. During the neural network training process, adaptive learning rate and mini-batch gradient descent methods were employed to enhance training efficiency. By embedding this surrogate model into the multi-physics iteration process, the computational efficiency of the coupled simulations was significantly improved. At the beginning, the surrogate model was used to determine the coupled physical field information for each node based on input parameters rapidly. Structural deformation calculations were performed according to the temperature distribution to obtain gap heat conduction information. Then, the thermal-hydraulic calculations and neutronics calculations were sequentially performed to update the temperature information, neutron flux density distribution, and power distribution. The fully connected neural network model and its accuracy in multi-physics coupling calculations were validated using the MMR core. The acceleration performance was thoroughly analyzed. Numerical results demonstrate that the neural network-based acceleration method achieves notable speedup without compromising computational accuracy.