基于神经网络的热管反应堆多物理场耦合快速预测

Rapid Prediction of Multi-physics Coupling in Heat Pipe ReactorBased on Neural Network

  • 摘要: 本文基于热管反应堆多物理场耦合分析框架,提出了一种基于数据驱动型神经网络构建的快速预测方案。该方案利用神经网络部分代替分析框架中的数值计算模块,实现核-热-力多物理场耦合迭代过程的高效求解。以Megapower作为分析对象,搭建了相应的神经网络模型构建快速预测方案并对其进行相关参数的预测,与传统数值计算方法对比。快速预测产生的关键参数和数值计算的结果差异很小,其中,最大应力差值不超过2 MPa,燃料平均温度差值不超过3 K。相同计算环境中,相比于数值计算需要6 h的计算时间,快速预测仅需4 min,计算时间降低了95%以上。基于上述结果,认为基于神经网络的快速预测方案具有准确性高、速度快的特点。配合其灵活性及对训练数据量要求低、可针对性的优化模型等特点,认为基于神经网络的快速预测方案是应对堆芯优化等大规模计算需求场景的优选方案。

     

    Abstract: In the present study, a data-driven neural network methodology was proposed for the rapid prediction of multi-physics interactions in heat pipe reactors, and grounded in a comprehensive framework of multi-physics coupling. By leveraging neural networks, certain numerical computation modules within the framework were effectively replaced, which facilitated an efficient resolution of the iterative process inherent to the nuclear-thermal-mechanical multi-physics coupling. In the process of multi-physics coupling, there exists a sophisticated iterative mechanism, wherein the computational analyses of distinct physical domains largely function autonomously. When contemplating the integration of neural network paradigms as surrogates within the multi-physics coupling framework, it is pivotal to target the numerical computations intrinsic to individual physical domains. By deploying one or multiple neural network models as substitutes, this methodology not only sustains the scalability inherent to rapid predictive schemes but also attenuates the complexity associated with the neural network’s surrogate functionalities. Concurrently, this approach necessitates a reduced dataset for training, and through the amalgamation of diverse neural network models, markedly augments the adaptability of the overarching rapid predictive framework. Furthermore, multiple neural networks, operating autonomously, can be dynamically calibrated with respect to their training data volumes based on their predictive performance. This not only optimizes their predictive accuracy but also facilitates the cost-effective expansion of training datasets in accordance with specific model requirements. Taking “Megapower” as the subject of analysis, pertinent neural network models were constructed and employed for rapid predictive parameter estimation, with the results juxtaposed against traditional numerical computations. The discrepancies between the key parameters yielded by rapid predictions and those from numerical simulations were minimal. Specifically, the maximum stress differential does not exceed 2 MPa, and the average fuel temperature difference is less than 3 K. Notably, while numerical computations demand six hours of processing time, the rapid prediction on the same platform requires a mere 4 minutes, marking a reduction in computational time by over 95%. Given these outcomes, it can be posited that neural network-based rapid prediction schemes exhibit high precision and accelerated processing speeds. Coupled with their inherent flexibility and modest training data requirements, and the capacity for targeted model optimization, it is advocated that neural network-based rapid prediction serves as a preferred methodology for addressing challenges such as core optimization in scenarios that demand large-scale computations. Concurrently, the neural network models utilized in rapid predictions have less stringent computational resource requirements and offer greater adaptability to varied environments. Capitalizing on these attributes, in conjunction with contemporary compact mobile reactors, one can facilitate on-site deployments with relative ease. Such integrations pave the way for prompt alerts in accident scenarios, thereby enhancing the safety of the reactor core.

     

/

返回文章
返回