2025-04-17 中国科学院（CAS）
Neural network model diagram(Image by CHEN Qisheng)
＜関連情報＞
- https://english.cas.cn/newsroom/research_news/phys/202504/t20250418_1041519.shtml
- https://www.sciencedirect.com/science/article/abs/pii/S0029549325001104?via%3Dihub
適応型ニューラルネットワークに基づく放射線遮蔽設計スキームの予測 Prediction of radiation shielding design schemes based on adaptive neural networks
Qisheng Chen, Zi-Hui Yang, Zhong-Yang Li, Guo-Min Sun, Shi-Peng Wang, Yu-Chen Li, Zhi-Xing Gu, Fei Li, Juan Fu, Gui-Hua Tao
Nuclear Engineering and Design Available online: 27 February 2025
DOI:https://doi.org/10.1016/j.nucengdes.2025.113933
Highlights
- This paper proposes a novel adaptive Deep Neural Network (DNN) model.
- This model can quickly generate a large number of radiation shielding design schemes and outperforms traditional models in terms of efficiency and speed.
- Verified by the space reactor model, the model demonstrates high accuracy and anti-interference ability.
- Its stability in complex environments is also guaranteed.
- In addition, the model has broad application potential in other radiation-related fields.
Abstract
As space exploration and space reactor technology continue to advance, radiation shielding design faces numerous challenges, such as space limitations, weight constraints, and the complexity of shielding materials. Traditional design methods typically rely on empirical models validated through Monte Carlo simulations, but these approaches often fail to achieve optimal results. To enhance the radiation protection efficiency of space reactors, this paper proposes a deep neural network model based on the self-attention mechanism to assist in predicting radiation shielding design schemes and to verify its accuracy and practicality. We used SuperMC software to record the relevant safety parameters for the Kilopower reactor under 10,000 different shielding design schemes and calculated the total mass and total radiation dose for these designs, creating a comprehensive dataset. The total mass and radiation dose were used as inputs to the neural network, which then generated the corresponding radiation shielding design schemes. Experimental results show that the model demonstrates high accuracy and strong interference resistance, with the error in total radiation dose and material mass consistently controlled around 3%. Additionally, by combining simulation methods with the self-attention mechanism, the model effectively generates radiation shielding designs suitable for space reactors, providing reliable protection solutions for future space missions. This approach also opens new possibilities for radiation shielding design in other fields.
