Abstract Comprehensive knowledge of neutron spectrum is significant in reactor design. Online wide-range neutron spectrum unfolding technology still requires improvement in accuracy and efficiency. In the work, a “two-step” neutron spectrum unfolding method based on artificial neural network (ANN) was developed to unfold spectrum with wide energy range. First, a default spectrum was reconstructed by using the ANN model which had been trained with a large amount of neutron spectra generated from Monte Carlo transport calculation. Second, the default spectrum was optimized by using iteration algorithm. The two-step method was verified with a thermal neutron reactor VENUS-3 and a fast neutron reactor BN-600. Comparison of mean square error (MSE) between this method and the traditional unfolding method showed reduction of 83.4% and 85.6% on VENUS-3 and BN-600 respectively, and average relative deviation (ARD) reduction of 89.3% and 86.1% respectively. Also, comparison of spectrum quality (Qs) showed reduction of 83.4% and 86.0% respectively for the two cases. This work demonstrated that the developed two-step method could obtain the better accuracy than traditional method.

中文翻译：

---

基于人工神经网络的裂变反应堆两步中子谱展开方法

摘要 中子能谱的综合知识在反应堆设计中具有重要意义。在线大范围中子能谱展开技术在精度和效率上还有待提高。在这项工作中，开发了一种基于人工神经网络（ANN）的“两步”中子谱展开方法，以展开具有宽能量范围的谱。首先，通过使用 ANN 模型重建默认谱，该模型已经用蒙特卡罗输运计算生成的大量中子谱进行了训练。其次，使用迭代算法优化默认光谱。两步法通过热中子反应堆 VENUS-3 和快中子反应堆 BN-600 进行了验证。该方法与传统展开方法的均方误差 (MSE) 比较显示减少了 83.4% 和 85。VENUS-3 和 BN-600 分别降低 6%，平均相对偏差 (ARD) 分别降低 89.3% 和 86.1%。此外，频谱质量 (Qs) 的比较表明，这两种情况分别降低了 83.4% 和 86.0%。这项工作表明，所开发的两步法可以获得比传统方法更好的精度。