Share

A neural-network approach to develop algebraic correlations for heat transfer and fluid flow.

Number: 2120

Author(s) : LIN L., GAO L., HWANG Y., KEDZIERSKI M.

Summary

Many heat transfer and fluid flow problems are too complex to model using traditional regression methods. Machine learning (ML) offers a new way to develop predictive models with high accuracy. However, current ML models are often uninterpretable and used as “black boxes”. This paper presents an approach to develop explicit, algebraic correlations from neural networks. An interpretable neural network, namely DimNet, is designed. One can train DimNet with experimental or simulation data and then convert the trained network to an explicit, power-law-like piecewise function. Besides being interpretable, DimNet inherits advantages of neural networks in modeling complex nonlinear problems. The mechanism and effectiveness of DimNet and the correlation development approach is further demonstrated by two case studies: 1) correlating simulation data for the friction factor of flow in smooth pipes; 2) correlating experimental data for flow boiling heat transfer coefficient within microfin tubes. Both case studies show DimNet can produce simple, explicit, algebraic correlations that are both statistically and phenomenologically accurate. The presented approach can be potentially used to develop correlations for various thermal-hydraulic problems, such as the pressure drop and heat transfer of single- and multi-phase flow, heat exchangers, and other thermal-hydraulic equipment.

Available documents

Format PDF

Pages: 10 p.

Available

Free

Download

Details

  • Original title: A neural-network approach to develop algebraic correlations for heat transfer and fluid flow.
  • Record ID : 30030473
  • Languages: English
  • Subject: Technology
  • Source: 2022 Purdue Conferences. 19th International Refrigeration and Air-Conditioning Conference at Purdue.
  • Publication date: 2022/07/10

Links


See other articles from the proceedings (224)
See the conference proceedings

Indexing