The free-return orbit serves as the preferred orbital scheme for crewed spacecraft in Earth-Moon transfers, yet its design involves stringent constraints and significant initial-value dependency in existing algorithms. This study addresses the Earth-Moon transfer trajectory planning for manned lunar exploration by proposing a dual-path neural network learning method to optimize free-return orbit initialization. First, a free-return orbit design framework is established, and the near-earth solution space characteristics are analyzed. Second, integrating the spatial partitioning characteristics of ascending and descending orbital phase in solution spaces, a dual-path neural network architecture designed via parameter-correlated transformation is proposed to ensure the completeness of orbital solutions. Finally, utilizing astromaster of Aerospace Tool Kit, the Earth-Moon free-return orbit planning under the dual-path network learning-based initialization method is implemented and validated through simulation. The results provide an effective reference for mitigating initial-value dependency in manned lunar mission orbit design.