Skip to main content
Nuclear Science and Engineering

AI/ML Uncertainty Quantification

Qualifying the accuracy of digital twin models

The use of a digital twin, whether for safety, monitoring or operations, is accompanied by a need to quantify the uncertainties inherent in the model. Uncertainty quantification (UQ) is a key element of nuclear reactor safety and is a requirement in the NRC’s advanced reactor licensing framework.  UQ enables reactor developers to reduce regulatory licensing effort and the cost of operations and maintenance.

There are two major types of UQ for digital twin models. Forward UQ can be used when the sources of uncertainty associated with the simulations are well-known, so that one can directly propagate all sources of uncertainty through the solver to obtain the uncertainty of simulation results. However, when we lack the knowledge of the sources of uncertainty in the simulation, inverse UQ is needed, which relies on new observations and machine learning methods to inversely quantify these sources of uncertainty before propagating them for simulations uncertainty.

In the inverse UQ approach, implementing a trained ML model in a physics-based code (such as SAM) introduces uncertainty, which can come from imperfect training data from either a CFD solver or measurement error or it can come from ML training error. These sources of uncertainty are not clearly known and can be estimated using the inverse UQ approach.

Three applications of inverse UQ related to Surrogate ModelsInherent Safety Response, and Feedwater Pump-Motor Set are described at the associated links.