Case study on sampling techniques using machine learning and simplified physical model for simulation-based dynamic probabilistic risk assessment

Kubo, Kotaro   ; Zheng, X. ; Ishikawa, Jun ; Sugiyama, Tomoyuki ; Jang, S.*; Takata, Takashi*; Yamaguchi, Akira*

Dynamic probabilistic risk assessment (PRA) enables a more realistic and detailed analysis than classical PRA. However, the trade-off for these improvements is the enormous computational cost associated with performing a large number of thermal-hydraulic (TH) analyses. In this study, based on machine learning (ML), we aim to reduce these costs by skipping the TH analysis. For the ML algorithm, we selected a support vector machine; we built it using a high-fidelity/high-cost detailed model and low-fidelity/low-cost simplified model. As a result, the computational costs could be reduced by approximately 80% without significantly decreasing the accuracy under the assumed conditions.