Abstract: The deterministic one-step whole-core heterogeneous calculation method represents a significant advancement in reactor physics calculations by employing more accurate physical models. This approach achieves higher computational accuracy and has been successfully applied in pressurized water reactors (PWRs). However, in fast neutron reactors, medium-mass nuclides exhibit strong elastic scattering resonance phenomena, resulting in highly oscillatory neutron flux distributions. To properly characterize these effects, the energy spectrum needs to be divided into thousands of groups for precise analysis. Performing full-core transport calculations with such ultra-fine energy group resolution using the deterministic one-step method proves computationally intensive. Conventional two-step approaches rely on equivalent or partitioned offline local models derived from the actual problem. In these methods, fine-group neutron spectra are computed offline to generate broad-group cross-sections, typically ranging from 30 to 50 groups. However, these offline-generated broad-group cross-sections cannot adequately represent spatial variations in cross-section characteristics across different core regions, leading to significant computational inaccuracies. To address the computational challenges of thousand-group transport calculations in fast reactor simulations using the deterministic one-step transport code SHARK, an online energy group condensation technique was developed to maintain both high accuracy and computational efficiency. For fuel regions, this method successfully establishes an online equivalent model based on escape cross-section conservation from resonance calculations, accounting for environmental effects. This enables online computation of thousand-group spectra and generation of broad-group cross-sections, representing significant progress in applying deterministic one-step methods to fast reactor analysis. Nevertheless, challenges remain in applying the online energy group condensation technique to non-fuel regions. Resonance calculation data are insufficient for constructing equivalent models in these areas, and offline-generated thousand-group spectra for broad-group cross-sections cannot adequately incorporate environmental effects. Moreover, the broad-group structures traditionally used in two-step methods for offline cross-section generation lack the precision required for online group condensation. Broad-group structures with excessive group numbers would drastically increase computational time, highlighting the need for optimized broad-group structures suitable for thousand-group condensation in deterministic one-step methods. To overcome these challenges, an online one-dimensional cylindrical model based on volume conservation using unstructured coarse meshes was implemented in non-fuel regions and their adjacent multi-layer coarse meshes was implemented to generate broad-group cross-sections. For the thousand-group structure, a multi-level particle swarm optimization approach was developed to obtain high-precision 104-group broad-group structures. Numerical results demonstrate that the optimized 104-group structure achieves high accuracy, with eigenvalue deviations within 100 pcm and power distribution deviations below 1%. This work advances the application of deterministic one-step methods in fast reactor analysis by balancing computational efficiency and accuracy.