To address the issues such as poor adaptability and insufficient generalization capability of traditional hanger force optimization methods for tie-arch bridges, this paper proposes a novel optimization framework integrating deep learning with multi-objective evolutionary algorithms. This method generates a "hanger force-structural response" dataset via a parametric finite element model. This method constructs a high-precision surrogate model using a residual connection-based neural network (ResNet), introduces the tuning framework based on Bayesian optimization to achieve automatic hyperparameter tuning of the surrogate model, and integrates the surrogate model with the Fast Elitist Multi-Objective Genetic Algorithm (NSGA-II) to achieve global optimization with bending strain energy, tie beam deflection, and hanger force uniformity as optimization objectives. Through the comparative analysis of examples, it is shown that the multi-objective optimization method proposed in this paper can significantly improve the internal force distribution of the structure while ensuring the smooth alignment of the bridge compared with the single-objective optimization strategy, the maximum positive and negative bending moments are reduced by 37.5% and 55.3% respectively, and the force distribution of the optimized hanger is more uniform. This study provides a new and effective approach for the intelligent and efficient multi-objective optimization design of complex bridge structures.