Uncertainty Quantification for Multi-Scale Reflood Tests Using Neural Network Based Surrogate Models

Markov Chain Monte Carlo (MCMC) was utilized in uncertainty estimation for thermal hydraulic system calculation by Heo et al. (2018). Based on the MCMC method with Bayes' theorem, they conducted data assimilation using 1D small scale tests, FEBA and refined the 32 important physical models and 5 boundary conditions. Subsequently, based on the calibrated parameter distributions, a posteriori distributions for large scale (FLECHT-SEASET) and multi-dimensional (PERICLES) tests were obtained. In this study, to reduce the computing demand for thermal hydraulic system calculation during the MCMC simulation, the machine learning was used to develop surrogate models for the complex system. The machine learning has been recently used in many fields of engineering. A key methodology of the machine learning is the neural network. The neural network is a useful modeling tool to solve a complex problem of multi-physics and multi-system via deep learning and deep network. In this study, the data assimilation methodology using machine learning models was suggested for the thermal hydraulic system to determine the uncertainties of the modeling parameters and the boundary conditions and uncertainties on the code simulation results. The neural network model provided an alternative solution with a tremendous reduction in the computational demand for the calculation. For the forward uncertainty propagation performed as a blind calculation, parameters' uncertainty bands were mapped through the computational model to assess the uncertainty bands of the calculation results for the PERICLES. The result shows that the adjusted distributions of the simulation output mostly cover the experimental data.