Abstract

Traditional light water reactor simulations are usually either high fidelity, requiring hundreds of node-hours, or low fidelity, requiring only seconds to run on a common workstation. In current research, it is desirable to combine the positive aspects of both of these simulation types while minimizing their associated negative costs. Because neural networks have shown significant success when applied to other fields, they could provide a means for combining these two classes of simulation. This paper describes a methodology for designing and training neural networks to predict normalized pin powers and $k_{eff}$ within a reflective two-dimensional pressurized water reactor assembly model. The developed methodology combines computer vision approaches, modular neural network approaches, and hyperparameter optimization methods to intelligently design novel network architectures. This methodology has been used to develop a novel new architecture, LatticeNet, which is capable of predicting pin-resolved powers and $k_{eff}$ at a high level of detail. The results produced by this novel architecture show the successful prediction of the target neutronics parameters under a variety of typical neutronics conditions, and they indicate a potential path forward for neural network–based model development.

摘要

传统的轻水反应堆仿真通常要么是高保真度（需要数百个节点小时），要么是低保真度（仅需数秒即可在普通工作站上运行）。在当前的研究中，理想的是将这两种模拟类型的积极方面结合起来，同时最大程度地降低其相关的负面成本。由于神经网络在应用于其他领域时已显示出巨大的成功，因此它们可以提供一种将这两类模拟结合在一起的方法。本文介绍了一种设计和训练神经网络的方法，以预测归一化的引脚功率和${ķ}_{ËFF}$反射式二维压水堆装配模型中。所开发的方法结合了计算机视觉方法，模块化神经网络方法和超参数优化方法，以智能地设计新颖的网络体系结构。该方法已用于开发新颖的新架构LatticeNet，该架构能够预测引脚解析的功率和${ķ}_{ËFF}$高度详细。这种新颖的体系结构产生的结果表明，在各种典型的中子学条件下，成功地预测了目标中子学参数，它们为基于神经网络的模型开发指明了一条潜在的途径。