A deep learning approach to estimate temperature and flow velocity distributions of wire-wrapped fuel assemblies

Computational fluid dynamic (CFD) of thermal-hydraulic has been widely used to investigate the design of sodium-cooled fast reactors (SFR). However, the specific models for wire-wrapped fuel assemblies in SFR always require complex procedures to set up and expensive computational costs to obtain final simulation results. In this study, we proposed a deep learning-based (DL-based) approach to estimate the temperature and flow velocity distributions of wire-wrapped fuel assemblies. The DL-based approach consists of an Auto-encoder model and Deep neural networks model, which are trained separately through a combination of unsupervised and supervised learning methods. The DL-based model was trained to take the input of external temperature probe values and corresponding boundary conditions and directly output the temperature and flow velocity distributions, bypassing the expensive CFD calculation. The presented DL-based approach can accurately and quickly estimate the temperature, transverse velocity, and axial velocity distributions with average errors of 0.00597%, 0.00506%, and 0.00745%, respectively. To our knowledge, this is the first study that demonstrates the feasibility of using DL as an alternative model to CFD to estimate thermal-hydraulic distributions accurately and quickly in SFR.