Machine learning assisted modeling of thermohydraulic correlations for heat exchangers with twisted tape inserts

TL;DR

This paper explores the use of machine learning algorithms to accurately model heat transfer correlations in heat exchangers with twisted tape inserts, demonstrating the superior performance of neural networks over other methods.

Contribution

The study applies and compares Polynomial Regression, Random Forest, and Neural Networks for modeling heat transfer correlations, highlighting the effectiveness of neural networks in this context.

Findings

ANN outperforms PR and RF in predicting heat transfer coefficients

ML models achieve high R^2 and low MSE, indicating good accuracy

Hyperparameter optimization enhances model generalizability

Abstract

This article presents the application of machine learning (ML) algorithms in modeling of the heat transfer correlations (e.g. Nusselt number and friction factor) for a heat exchanger with twisted tape inserts. The experimental data for the heat exchanger at different Reynolds numbers and twist ratios were used for the correlation modeling. Three machine learning algorithms: Polynomial Regression (PR), Random Forest (RF), and Artificial Neural Network (ANN) were used in the data-driven surrogate modeling. The hyperparameters of the ML models are carefully optimized to ensure generalizability. The performance parameters (e. g. $R^2$ and $MSE$) of different ML algorithms are analyzed. It was observed that the ANN predictions of heat transfer coefficients outperform the predictions of PR and RF across different test datasets. Based on our analysis we make recommendations for future data-driven modeling efforts of heat transfer correlations and similar studies.