The design and safety analysis of advanced nuclear reactors often involves predicting complex thermal-fluid phenomena that occur in various reactor components. Accurately predicting such phenomena for long transients remains a key challenge for reactor safety analysis. This project aims to develop a hierarchical, multi-scale methodology that employs scientific machine learning (ML) techniques to improve safety modeling of advanced nuclear reactors in terms of both efficiency and accuracy. Physics-informed ML techniques has been developed and applied to existing simulation capabilities to create a computationally efficient, multiscale analysis framework for nuclear reactor safety analysis. This framework facilitates the development and deployment of advanced reactors by improving economics (through accurate safety margin predictions) and reducing the licensing burden (through improved uncertainty quantification).

This project involves development, optimization, and uncertainty quantification of physics-informed ML methods to improve the predictive capability for reactor safety modeling. The project team used multiscale, high-fidelity, computational fluid dynamics (CFD) simulation  as the database to train the machine learning model. Reynolds Averaged Navier Stokes (RANS) simulations are used to generate the full transient simulation results of reference cases, which are directly used in training the data-driven closure relations. Higher-fidelity large eddy simulations (LES) are used to simulate a few selected ​“snapshot” periods during the transients, which provide the reference solutions for uncertainty quantification of the ML model. State-of-the-art deep neural network architectures were used for baseline model development; optimizations were completed for both hyperparameters and neural network architecture. The team studied the uncertainty quantification of the ML model using three approaches: Monte Carlo dropout, deep ensemble, and Bayesian neural network.