Development of a data-driven simulation framework using physics-informed neural network

Abstract

Because nuclear power plants (NPPs) are safety-critical systems with large sizes and high complexities, various methods have been developed to identify possible accidents and address potential risks. Accordingly, the probabilistic safety assessment (PSA) methodology is a prevalent nuclear risk assessment methodology. However, PSA has limitation in the consideration of dynamic interactions. As simulation methodologies develop, several simulation-based D-PSA methodologies have been developed, but there is a problem that the calculation speed and interaction of simulation cannot be considered.We propose a physics informed neural network (PINN) based data-driven simulation framework. The simulation framework comprises solution and model generators. The solution generator performs as the existing physical process model by using initial conditions, and boundary conditions. The model generator configures the most representative equation by using measurement data. Therefore, by sing the proposed framework, it is possible to construct a simulation methodology that not only has a fast simulation speed, but can also easily reflect system interactions.