In this study, an elastoplastic constitutive model was developed for porous materials by combining plasticity theory with neural networks, aiming to conduct high-fidelity finite element simulations of the ductile fracture behavior in metallic materials. Compared to traditional machine learning methods, this approach reduces the amount of data required for model training, thus improving modeling efficiency. By using neural networks to learn the numerical data for the yielding behaviour of porous materials generated by representative volume element (RVE) simulations, a neural network-based yield surface for porous materials was constructed. This yield surface was compared with the Gurson model. Subsequently, the neural network-based yield surface was integrated into the rate-independent plasticity framework, in combination with a pore evolution law, and implmented in Abaqus through a UMAT subroutine using the radial return algorithm. Tensile tests on notched specimens were simulated numerically using the neural network-based constitutive model. Strain-stress curves and the evolution of void volume fraction were analyzed, as well as the equivalent plastic strain distribution. The results obtained by the simulations were in line with the experiments. This study not only provides a precise predictive tool for ductile fracture problems but also demonstrates high adaptability, laying a foundation for studying materials with more complex mechanical properties, such as the anisotropic behavior of additively manufactured titanium alloys.