[2C09-14] AI

3:15 PM - 3:30 PM JST(6:15 AM - 6:30 AM UTC)

[2C11] Prediction of Breakthrough Curves Using Deep Learning Informed by Fracture Network Structure and Fracture Properties

*Shunichi Okamoto^1,2, Kotaro Nakata¹, Koji Mori³, Takumi Saito² (1. CRIEPI, 2. UTokyo, 3. BES)

Keywords:

Fracture networks,Graph neural network,Solute transport simulation,Breakthrough curves,Latin hypercube sampling

Understanding radionuclide migration through fractured rock is essential to the safety assessment of geological disposal. Discrete fracture network (DFN) models are widely used to simulate solute transport in fractured media, with the results commonly summarized as breakthrough curves (BTCs). However, DFN generation parameters involve uncertainties due to limited site observation data, and evaluating their effects requires sensitivity analysis. Because transport simulations for large numbers of DFN realizations are computationally expensive, an efficient prediction method for BTCs is needed. In this study, we propose a BTC prediction framework that combines an efficient sampling strategy for DFN generation parameters with a graph neural network that learns fracture network structures and physical properties. The overall framework and predictive accuracy will be presented.