A general model for flow boiling heat transfer in microfin tubes based on a new neural network architecture

Published

March 19, 2022

Author(s)

Lingnan Lin, Lei Gao, Mark A. Kedzierski, Yunho Hwang

Abstract

A new neural network architecture, namely DimNet, was designed for correlating dimensionless quantities with power-law-like relations. Unlike common neural networks that are usually used as "black-boxes", DimNet is interpretable as it can be converted to an explicit algebraic piecewise power-law-like function. With DimNet, a data-driven, empirical model was developed to predict the pre-dryout heat transfer coefficient of flow boiling within microfin tubes. The model was trained on a database with 7349 experimental data points for 16 refrigerants, and then optimized by comparing different sets of dominant dimensionless quantities and by adjusting the network configuration. The model exhibits an overall mean-absolute-error of 13.8% and no systematic variation with respect to the salient parameters for most conditions. Besides being statistically accurate, the model captures parametric trends of the heat transfer coefficient. The excellent prediction performance of the model was attributed to the DimNet's ability to automatically classify the data into optimal regions and simultaneously correlate the data of each region. Therefore, the DimNet architecture is inherently suitable for modeling complex heat transfer and flow problems where multiple distinct physical regimes exist, especially for problems where a power-law-like input–output relation is desired such as convective heat transfer.

Citation

Energy and AI

Pub Type

Journals

Download Paper

https://doi.org/10.1016/j.egyai.2022.100151

Local Download

Keywords

convective boiling, micro-fin, two-phase flow, machine learning, ANN, dimensionless number

Citation

Lin, L. , Gao, L. , Kedzierski, M. and Hwang, Y. (2022), A general model for flow boiling heat transfer in microfin tubes based on a new neural network architecture, Energy and AI, [online], https://doi.org/10.1016/j.egyai.2022.100151, https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=932284 (Accessed December 14, 2024)