Uncertainty analysis of fault scenarios in geological disposal using neural networks

Abstract

  In this study, a realistic radionuclide transport model was developed by assuming the fault zone to consist of distinct sub-regions: the gouge zone, the damage zone, and the fracture zone, and by taking into account the specific characteristics of each region. Numerical simulations were conducted using this model, and the results were employed as training data to construct a predictive model through machine learning of the neural network. This predictive model enables fault impact assessments across the entire repository. Furthermore, focusing on major uncertainty factors related to fault scenarios, specifically fault size, occurrence timing, and location, a total of 1,000 uncertainty analysis cases were performed using various combinations of these factors and the developed predictive model. The results revealed that fault magnitude exerts the greatest influence on the maximum total dose, and that the dominant contributing region to the maximum total dose varies depending on the conditions. These findings suggest that, for an appropriate evaluation of the maximum total dose, it is essential to consider the number of waste packages located within each fault zone region.