Session: K8-01: FUNDAMENTALS OF MACHINE LEARNING IN HEAT TRANSFER

Paper Number: 131027

131027 - Prediction of Critical Heat Flux for Liquid Helium Using Machine Learning Models Assisted by Physics-Based Correlation 

Abstract: 

Cryogenic fluid management (CFM) is highly crucial for NASA’s space missions and one challenge is the tank fill process for cryogenic fluids. To date, no predictive tool can accurately estimate the pool boiling curve during this process. We aim to address this gap first from prediction of critical heat flux (CHF). We use both traditional machine learning (ML) methods and novel physics-assisted machine learning (PAML) methods first proposed by this study to predict CHF on a large database of 1322 datapoints from 52 published works with a wide range of operational conditions and thermophysical properties of fluids. While the ML approach used no prior knowledge in modeling, the PAML approach assumed the correlation form on Zuber’s hydrodynamic instability model [1] and predicted the multiplier ‘K’ to estimate the CHF.  Backward elimination, absolute shrinkage selection operator and univariate selection based on feature importance are employed to reduce the redundant features while preserving as much information as possible. Multiple linear regression, k-nearest neighbors, random forest (RF), extreme gradient boost (Xgboost) and artificial neural network (ANN) are used for modeling. Moreover, the feature importance of both ML and PAML models are presented based on the magnitude of disturbing error of each input feature. The major outcomes are (i) the RF, Xgboost and ANN for the traditional ML models resulted in similar performance compared to the traditional semi-empirical correlations, (ii) the novel PAML methods predicted much better than traditional ML methods for up to 17% of improvement on the best performed model, (iii) PAML models have much lower level of susceptibilities (error by disturbance and noise of data) than traditional ML models. The best predictions are obtained using Xgboost on PAML considering thermophysical properties of fluids resulting in a mean absolute percentage error of 12.79% which is far better than the semi-empirical correlations (21%-50%), providing the most accurate tool to date in estimating CHF for cryogenic fluids. This work is laying the groundwork for developing ML models that can predict the full pool boiling curve with high accuracy.

[1] Zuber, Novak. "On the stability of boiling heat transfer." Transactions of the ASME, 80.3 (1958): 711-714.

Presenting Author: Jiayuan Li Case Western Reserve University

Presenting Author Biography: The presenting author is a Phd student in Case Western Reserve University. He works in Dr. Chirag Kharangate's 'two phase thermal management lab' and mainly works on areas related to fundamentals of two phase heat transfer.

Authors:

Jiayuan Li Case Western Reserve University
Chirag Kharangate Case Western Reserve University