Session: Artificial Intelligence and Machine Learning Models

Paper Number: 152505

152505 - Critical Heat Flux Prediction and Uncertainty Quantification With Bayesian Optimization and Deep Ensemble 

Abstract: 

Understanding Critical Heat Flux (CHF) is fundamental for the safe design and operation of water-cooled nuclear reactors. CHF represents the point at which the heat transfer from a heated surface dramatically diminishes, causing a rapid rise in wall temperature. This phenomenon leads to either dryout in boiling water reactors (BWRs) or Departure from Nucleate Boiling (DNB) in pressurized water reactors (PWRs). The prediction of CHF is analytically challenging due to the complex interplay of local thermal and fluid dynamics, and as such, often relies on empirical correlations or look-up tables derived from experimental data.

Recent advancements in data-driven modeling and machine learning (ML) have provided new opportunities for improving CHF prediction accuracy. In this context, the Nuclear Energy Agency (NEA) launched an initiative to establish a benchmark for ML-based CHF predictions, motivated by the U.S. Nuclear Regulatory Commission (USNRC). The USNRC developed the Groeneveld CHF lookup table and released the NRC CHF database, which contains approximately 25,000 data points. As part of this benchmarking effort, we submitted an ML model employing five input features—outlet quality, pressure, mass flux, heated length, and tube diameter—to predict CHF. However, discrepancies were observed between the original CHF data and the model predictions.

The experimental data in the CHF database is known to be noisy, influenced by various sources of uncertainty. Given that the data points originate from multiple experiments, it is crucial to estimate the uncertainty associated with this noise—referred to as aleatoric uncertainty—to understand the limitations of any neural network applied to this data. Additionally, epistemic uncertainty, which arises from the model's limited knowledge, must also be considered. The sensitivity of a neural network to its hyperparameters, such as learning rate, activation function, dropout rate, and the number of hidden layers, contributes to epistemic uncertainty, as it reflects the limitations in model knowledge and the potential for enhancing predictions through better hyperparameter tuning. Therefore, optimizing the neural network is essential for accurate uncertainty quantification.

In this work, we utilize the deep ensemble method and Bayesian Optimization (BO) to optimize a deep neural network (DNN) trained on the NRC CHF database, focusing on estimating both aleatoric and epistemic uncertainties. This combined approach, referred to as BODE, helps improve the model's predictive capability and quantify uncertainties more effectively. To further enhance prediction performance, we augmented the input features by incorporating derived parameters such as inlet subcooling and Reynolds number. This optimized setup, achieved through Bayesian Optimization, demonstrated significantly improved performance compared to the baseline five-input feature model, resulting in reduced aleatoric and epistemic uncertainties.

Presenting Author: Zaid Abulawi Texas A&M University

Presenting Author Biography: Ph.D. Nuclear Engineering (Expected February 2028)
Texas A&M University, Texas
Advisor: Dr. Yang Liu

M.Sc. in SAfe and REliable Nuclear Applications (SARENA) program (September 2023)
Lahti-Lappeenranta University of Technology (LUT), Finland Mar 2023 – Sep 2023
University of Ljubljana (UL), Slovenia Oct 2022 – Mar 2023
Lahti-Lappeenranta University of Technology (LUT), Finland Jan 2022 – May 2022
IMT Atlantique, France Sep 2021 – Jan 2022
GPA: 9.5/10

B.S.E. Nuclear Engineering (July 2020)
Jordan University of Science and Technology (JUST), Jordan ABET accreditation
GPA: 3.78/4

Graduate Research Assistant
Nuclear Engineering Department, Texas A&M University Feb 2024 – present
- Applying machine learning techniques on nuclear reactor safety applications. The research includes
combining numerical methods, automatic differentiation, and machine learning, focused on thermalhydraulics and fluid flow problems.

Authors:

Zaid Abulawi Texas A&M University
Doyeong Lim Texas A&M University
Yassin Hassan Texas A&M University
Yang Liu Texas A&M University