F-MNet: flux-modulated network for radial power prediction of space reactors

Rapid and accurate prediction of radial power distribution is crucial for the design and operation of space nuclear reactors. However, traditional Monte Carlo methods incur prohibitively high computational costs, while conventional data-driven surrogate models face severe overfitting and insufficient generalization capabilities under small-sample and data distribution shift conditions. Addressing the problem of radial power prediction for rod-controlled space reactors, this paper proposes F-MNet (Flux-Modulated Network), a physics-informed neural network architecture. F-MNet employs the Feature-wise Linear Modulation (FiLM) mechanism to take the core-average neutron flux as a conditioning signal and dynamically adjust hidden-layer feature transformations, thereby learning the nonlinear sensitivity function of local assembly power with respect to flux—a relationship intrinsically constrained by the control rod differential worth. To evaluate the extrapolation performance under small-sample conditions, four dataset partitions with distinct characteristics of control rod position distributions were constructed. Ablation experiments demonstrate that F-MNet consistently outperforms all baseline models (Plain MLP, Res + FiLM, Res Only) across all partition schemes. Notably, in the challenging extrapolation scenario Case 1, F-MNet increased the R 2 from 0.4540 for Plain MLP to 0.9429, with the MAE, RMSE, and MAPE all reduced to approximately one-third. Furthermore, a cross-criticality case experiment was conducted, spanning subcritical (keff ≈ 0.62) to supercritical (keff ≈ 1.05) states, where the maximum relative deviation for all assemblies remained below 1.55%, further validating the engineering applicability of the model. The aforementioned results indicate that F-MNet is an efficient, accurate, and physically consistent model framework, providing a reference design paradigm for surrogate models dealing with small-sample and data distribution shift scenarios.