Performance evaluation of a surrogate model to predict effective dose under hypothesized severe accidents of pressurized water reactor based on neural network

Abstract

Numerous studies utilizing severe accident analysis codes have been conducted to evaluate the effectiveness of various mitigation strategy combinations. However, due to various sensitivity factors such as the execution time and characteristics of system performance, an exhaustive analysis of all scenarios is infeasible. Interestingly, recent studies have shown that application of the machine learning technique is beneficial for reducing the calculation cost and this study developed a deep neural network-based surrogate model to assess the effectiveness of a severe accident management strategy. The effectiveness of the strategy was evaluated based on the effective dose over a 72 hrs, which is an ultimate standard for measuring the ability to mitigate a severe accident. Additionally, the model's performance was further improved by incorporating input parameters that can capture the progression of accident, such as the timing of major events and accident classification based on event branches using a source term grouping method.