Deep learning-based visual prediction of hydrogen distribution in a passive autocatalytic recombiner

Abstract

Maintaining the operational stability of Nuclear Power Plants (NPPs) during severe accident environments requires a comprehensive understanding of local hydrogen distribution. To mitigate hydrogen-associated risks, Passive Autocatalytic Recombiners (PARs), which operate passively through catalytic reactions, are extensively installed within the containment structures. This study numerically investigates the thermohydraulic behavior of PARs under various accidental conditions, specifically focusing on the effects of inlet velocity and temperature. Building upon previous research (Kim et al. [14]), this work aims to predict hydrogen distribution within PARs using only temperature and velocity data, applying Convolutional Neural Networks (CNN) with a U-Net and Auto-Encoder (AE) models featuring different encoder-decoder architectures (MLP-CNN, MLP-ResNet). The U-Net model shows outstanding predictive performance, achieving a MAPE of 2.32 %, MAE of 0.042, RMSE of 0.080, and Structural Similarity Index Measure (SSIM) of 0.99, using temperature and velocity contours as input data. AE models, which utilize low-dimensional thermohydraulic inputs such as inlet velocity, inlet temperature, and outlet temperature, also exhibit strong predictive capability. The MLP-CNN and MLP-ResNet architecture exhibit reliable performance, with MAPE below 2.8 %, MAE below 0.52, RMSE below 0.09, and SSIM above 0.92. This research highlights the feasibility of effectively visualizing hydrogen distribution in PARs through advanced data-driven models.