A method fusing self-attention-SPN and NSGA-III for multi-objective radiation shielding design optimization

Abstract

High-performance compact nuclear facilities require radiation shielding designs that balance safety and weight considerations. Current intelligent shielding design methods typically combine Evolutionary Algorithms (EA) for optimization with neural networks for evaluation. However, the neural networks used, primarily BP or DNN models composed of Fully Connected (FC) layers, require large datasets and extensive computation resources. A novel method fusing Self-Attention-based Sequence Prediction Network (Self-Attention-SPN) and Non-dominated Sorting Genetic Algorithm III (NSGA-III) was proposed in this paper for multi-objective radiation shielding design optimization. By reformulating dose rate calculation as a sequence prediction problem, the SPN of lightweight network structure leverages the multi-physics feature projection and multi-head self-attention mechanism to effectively capture the inter-layer physical feature relationships, ensuring high prediction accuracy with small datasets. The method is validated using the Savannah reactor case, where SPN achieves Monte Carlo (MC)-level accuracy with significantly reduced computational cost. Comparative experiments show that training data with additional physical parameters can reduce SPN training loss, underscoring the importance of physical information. Furthermore, SPN outperforms BP in prediction accuracy, validating the effectiveness of the multi-head self-attention mechanism. Sensitivity analysis of NSGA-III coupled with SPN prediction perturbation confirms the robustness of the proposed method. The optimization solutions effectively converge to the Pareto front, demonstrating the method’s efficiency and reliability for multi-objective radiation shielding design.