Abstract: To address the problem of interfacial parameter determination in the meso-scale damage simulation of steel fiber reinforced concrete (SFRC), this study collected and analyzed data from 189 single fiber pullout tests to develop deep neural network-based models for predicting the mechanical response of the bonding interface during fiber pullout. Their optimal hyperparameters were tuned via the Bayesian optimization algorithm combined with 5-fold cross-validation. Furthermore, an efficient 3D three-phase SFRC damage analysis method was proposed to explicitly simulate the matrix, fibers, and fiber-matrix interface, with the bond-slip constitutive parameters of the interfacial elements derived directly from these prediction models. The developed interfacial parameter prediction models demonstrate excellent generalization capability, achieving R² values exceeding 0.95 on the training set and 0.93 on the test set. The proposed damage analysis method requires only 1.51 and 13.40 minutes to solve the single fiber pullout and three-point bending beam examples, respectively, marking time reductions of 99.7% and 89.9% compared to existing benchmark methods. The results show that the proposed method overcomes the limitations of the current approaches for SFRC interfacial constitutive parameter determination, and effectively solves the persistent problems in SFRC numerical methods, such as complex mesh generation, ambiguous interface parameters, poor convergence, and high computational cost. Thus, it provides a novel solution for the efficient damage analysis of SFRC structures at the engineering scale.