Anil Gurgen
Department of Nuclear Engineering
North Carolina State University

Abstract

Sensor reliability is critical for safe and economic operation of nuclear plants. Current sensor calibration methods are labor intensive, expensive and introduce potential risk of additional fault. Online sensor calibration monitoring using data analytics and machine learning techniques can avoid these issues by identifying which sensor need calibration and this allow plant engineers to extend sensor calibration cycle. Machine learning techniques are capable of predicting sensor response based on implicit relationships between the measurements from multiple sensors. These techniques are involved as regression tools, and quantifying the uncertainty associated with the model prediction is an important step in online sensor fault detection. Generalization is a measure on how model will perform against the new data and in this talk, the speaker will introduce the fundamentals of generalization theory for machine learning regression algorithms and present  an analytic method to estimate generalization equations for two sample regression algorithms; support vector machines regression and auto-associative neural networks. A brief discussion about the implications of these equations will follow.

Biography

Anil Gurgen is a Ph.D. student in Nuclear Engineering Department at North Carolina State University. He got his bachelor degree of nuclear engineering in 2013 from Hacettepe University, Turkey and master degree of nuclear engineering in 2018 from MIT. He is currently involved in the ARPA-E project “Development of a nearly autonomous management and control system for advanced reactors” and doing his research under the supervision of Dr. Dinh. His research mainly focuses on identifying requirements and uncertainties in machine learning algorithms that can be used to predict future states of reactor in transient conditions.

Joseph Coale
Department of Nuclear Engineering
North Carolina State University

Abstract

 A parametric study was performed on a LANL rad-flow experimental campaign known as the COAX diagnostic to understand better uncertainties. COAX uses a half-hohlraum radiation drive to push a supersonic wave of radiation through a foam target, and current simulations assume the radiation drive is an isotropic and black-body source of radiation. The geometry and material make-up of the half-hohlraum motivate investigation into the effect of an anisotropic drive with a non-planckian frequency distribution, which may change how the supersonic radiation front propagates through the target. The performed study prompted further investigation to tie these effects into commonly used opacity scalings, and formed a basis for another parametric study using reduced-order models to enable much more data to be gathered in a short period of time.

Biography

Joseph Coale is a Ph.D. student in the Nuclear Engineering Department at North Carolina State University. He got his bachelor degree of nuclear engineering in 2018 from North Carolina State University. He currently is doing his research under Dr. Dmitriy Anistratov and his thesis work is on reduced-order modelling for thermal radiative transport problems.

Ryan O’Mara
Department of Nuclear Engineering
North Carolina State University

Abstract

Sampler is a module within the SCALE code package that is designed for total uncertainty quantification in radiation transport problems [5]. Sampler provides an automated way to stochastically sample the possible nuclear data and nominal physical parameters in a simulation and then run the perturbed cases and collect the results. Additionally, Sampler can be combined with any type of SCALE calculation to allow for comprehensive uncertainty quantification for even very complex radiation transport problems. Recently, efforts have begun to implement kernel density estimations into the Sampler architecture in an effort to better characterize the total propagated uncertainties in Sampler analyses. densities. Kernel density estimators (KDEs) are a means of including sample values and the uncertainty associated with those samples in order to better characterize the underlying population density. This talk will cover the basics of KDE analysis and the efforts to include KDEs as part of the standard analysis supported by Sampler.

Biography

Ryan is a PhD candidate studying retrospective dosimetry applications to nonproliferation and nuclear fuel characterization, under the advisement of Dr. Hayes. This past summer Ryan held an internship at Oak Ridge National Lab. There he worked with the SCALE team on adding new analysis capabilities for stochastic uncertainty propagation studies. Part of this work included adding kernel density estimator analysis to the Sampler package in SCALE.

***This seminar will be streamed live on our NCStateNuclear YouTube channel***