Anil Gurgen
Candidate for PhD in Nuclear Engineering

Development and Assessment of Physics-guided Machine Learning Framework for Prognosis System

Monday, April 26. 2021
10:00 am – 12:00 pm

https://ncsu.zoom.us/j/98056300742

Nearly Autonomous Management and Control (NAMAC) is a system that provides recommendations to help operators in decision-making during plant operations ranging from normal operation to accident management. A critical function in NAMAC is prognosis. In nuclear engineering domain, prognosis is the process of predicting future conditions of a system or equipment based on the present signs and symptoms of a fault. The prognosis model is an integral part of autonomous decision-making and NAMAC by predicting future reactor states for possible candidate control strategies so that the outcomes can be evaluated to determine the best control strategy. The prognosis model requires representing direct relationships between the symptoms and the predictions. The traditional approach for representing such relationships is using a model-based approach where models represent scientific knowledge. In nuclear engineering, computational simulations are approximate representations of the operation of the real system. However, prognosis with computational simulations requires high computation power and time due to the possibly large number of scenarios, and this requirement contradicts the faster-than-real-time decision-making philosophy of NAMAC. Necessary computation resources can be reduced with a machine learning (ML) approach for fast predictions by building a surrogate function using the simulation data. A critical issue is, ML models are ignorant of physical knowledge, and these models approximate statistical relationships between the system variables. This ignorance can produce results that are inconsistent with physical laws, even if an optimal result is achieved from a mathematical point of view. To avoid such problems, physics-guided machine learning (PGML) is proposed in the ML community.

This dissertation has two primary objectives. The first objective is to identify a framework to guide development, assessment, and uncertainty quantification of the ML-based prognosis model. Training of a ML model consists of optimizing many aspects of the ML approach. Also, the predictions of the prognosis model are crucial for NAMAC’s decision-making process; therefore, assessment of accuracy and quantification of uncertainty of the ML-based prognosis model is an essential part of NAMAC development. For the first objective, an optimization framework is formulated to suggest an optimal surrogate function for the prognosis model and provide testing conditions to assess the accuracy and uncertainty of the prognosis model. The second objective of the dissertation is to identify the requirements of PGML in order to improve the ML-based prognosis model. For the second objective, the PGML approach is reviewed, and a physics-guide model is developed and used in the training of the ML model. To achieve the objectives, four prognosis models are developed with the optimization framework to predict fuel centerline temperature based on reactor state variables by training Recurrent Neural Network for a loss-of-flow accident in Experimental Breeder Reactor II. The developed prognosis models are pure data-driven prognosis (PDP), regularized data-driven prognosis (RDP), physics-guided data-driven prognosis (PGP), and regularized physics-guided data-driven prognosis (RPGP), with different contributors in the loss function of ML training. Prognosis models are successfully developed with the optimization framework, and the assessment of prognosis models provides insights on the accuracy and uncertainty of prognosis models. Based on the insights gained, prognosis models are further evaluated considering the NAMAC system requirements. The results indicate that PGP has the best performance compared to other prognosis models.