The use of machine learning for inverse uncertainty quantification in TRACE code based on Marviken experiment

Domitr, Paweł; Włostowski, Mateusz

doi:10.1016/j.nucengdes.2021.111498

Cited by 3 publications

(1 citation statement)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The inverse UQ is usually performed using the Bayesian framework due to its robustness in parameter inference as demonstrated by the statistics [2] and machine learning [3] areas. The value of inverse UQ can already be found in many disciplines, including nuclear engineering [4,5] to calibrate nuclear thermal-hydraulics codes, in aerospace [6] to characterize complex damage scenarios of aerospace components, in nuclear fuel materials [7] using variational Bayesian, in composite science [8] to quantify uncertainties of a carbon fiber reinforced composite, in multiphase computational fluid dynamics [9], and many other examples. For the efforts focusing on the theory, the seminal work by [10] introduced the Bayesian inverse framework to calibrate computer models, which has been used as the major building block for the next inverse UQ efforts.…”

Section: Introductionmentioning

confidence: 99%

Bayesian Inverse Uncertainty Quantification of the Physical Model Parameters for the Spallation Neutron Source First Target Station

Radaideh,

Lin,

Jiang

et al. 2022

Preprint

View full text Add to dashboard Cite

The reliability of the mercury spallation target is mission-critical for the neutron science program of the spallation neutron source at the Oak Ridge National Laboratory. We present an inverse uncertainty quantification (UQ) study using the Bayesian framework for the mercury equation of state model parameters, with the assistance of polynomial chaos expansion surrogate models. By leveraging high-fidelity structural mechanics simulations and real measured strain data, the inverse UQ results reveal a maximum-a-posteriori estimate, mean, and standard deviation of 6.5 × 10 4 (6.49 × 10 4 ± 2.39 × 10 3 ) Pa for the tensile cutoff threshold, 12112.1 (12111.8 ± 14.9) kg/m 3 for the mercury density, and 1850.4 (1849.7 ± 5.3) m/s for the mercury speed of sound. These values do not necessarily represent the nominal mercury physical properties, but the ones that fit the strain data and the solid mechanics model we have used, and can be explained by three reasons. First, the limitations of the computer model or what is known as the "model-form uncertainty" that would result from numerical methods and physical approximations. Second, is the biases and errors in the experimental data. Third, is the mercury cavitation damage that also contributes to the change in mercury behavior. Consequently, the equation of state model parameters try to compensate for these effects to improve fitness to the data. The mercury target simulations using the updated parametric values result in an excellent agreement with 88% average accuracy compared to experimental data, 6% average increase compared to reference parameters, with some sensors experiencing an increase of more than 25%. With a more accurate strain response predicted by the calibrated simulations, the component fatigue analysis can utilize the comprehensive strain history data to evaluate the target vessel's lifetime closer to its real limit, saving tremendous target cost and improving the design of future targets as well.

show abstract