pcNN-SN couples physics-enhanced networks to SN transport sweeps, cutting 3D neutron solve time over 60 percent while preserving conservation

{"The algorithm trains on paired high- and low-accuracy SN solutions, feeding the network dimensionless subgrid-scale features extracted directly from the transport equation. Physical constraints are enforced so that the corrected flux retains particle balance. Internal coupling allows the network to supply online correction factors during the discrete-ordinates sweeps rather than post-processing.","Numerical experiments show the corrected coarse-mesh results match fine-mesh reference solutions within discretization error for both k-effective and flux distributions. Reusability tests across differing cross-section sets indicate the physics-based input transformation narrows the parameter distribution enough for the same trained model to generalize without retraining.","Traditional mesh-refinement studies require proportional growth in memory and sweep time; the pcNN-SN approach decouples accuracy from mesh density. This directly addresses the bottleneck in full-core reactor analysis where fine angular and spatial resolution remains prohibitive. The 60 percent time saving reported for the 3D case suggests potential for routine use in uncertainty-quantification loops that currently rely on coarse models.","Next steps include extension to time-dependent problems and coupling with Monte-Carlo tallies. Validation on full-scale pressurized-water-reactor geometries with thermal-hydraulic feedback will determine whether the observed reusability persists under realistic material heterogeneities."}