A MACHINE LEARNING APPROACH FOR THE PREDICTION OF FLOW BOILING HEAT TRANSFER COEFFICIENTS IN SMALL TO MICRO-TUBES

Aug 6, 2026

Flow boiling in microchannels is central to next-generation cooling systems, yet designing efficient devices requires understanding tightly coupled microscale phenomena that are difficult to represent with physics-based equations. This study applies machine learning regression to predict heat transfer coefficients in single small-to-micro tubes (0.52–4.26 mm) using variables extracted from the Brunel Two-phase Flow Database. Results indicate accurate predictions within training data ranges, while extrapolation needs caution.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages3

Reading Time8 min

AccessFree to download

Abstract

Introduction

Methodology

Results

Data preprocessing and feature selection

Neural network training and evaluation

Conclusions

A MACHINE LEARNING APPROACH FOR THE PREDICTION OF FLOW BOILING HEAT TRANSFER COEFFICIENTS IN SMALL TO MICRO-TUBES Nima Nazemzadeh 1, Francesco Coletti 1,2*, Tassos G. Karayiannis2 1Hexxcell Ltd., Foundry Building, 77 Fulham Palace Road, W6 2AF, London, United Kingdom 2Brunel University London, Uxbridge, Middlesex UB8 3PH, United Kingdom 1. ABSTRACT Flow boiling in microchannels plays an important role in the future of cooling systems. However, to design efficient devices that exploit the benefits of latent heat cooling, it is necessary to develop a detailed understanding of the several complex phenomena that interact with each other and are extremely challenging to capture mathematically with physics-based models. This study explores the application of machine learning (ML) algorithms to demonstrate their predictive abilities in the absence of detailed deterministic knowledge. The work leverages the extensive Brunel Two-phase Flow Database to extract the explanatory variables needed for predictions and uses various regression models to predict the heat transfer coefficient in single small to micro tubes with diameters ranging from 0.52 to 4.26 mm. The preliminary results demonstrate that the ML algorithm can predict accurately, albeit caution is needed when extrapolating beyond the ranges of the data used for training. 2. INTRODUCTION The application of micro-evaporators for cooling electronic devices has the potential to reduce the capital cost when manufacturing thermal management systems used in a number of industries and applications including renewable energy systems, electric vehicles and charging stations, refrigeration and components in computers and information technology systems and other high-power semiconductor devices, [1] . However, designing such systems remains a challenge due to the complexity of the phenomena involved at the microscale. Several research studies have focussed on developing mechanistic or statistically derived models to predict the heat transfer coefficient (HTC) in such systems [2–4] . The desired output is a simple yet accurate way to predict the heat transfer coefficient needed for the design. This is currently a developing area and it is now considered that machine learning algorithms could also have a role to play and should be considered as they have the ability to estimate the desired output variables (e.g. the HTC) by learning the underlying complex correlations among the input variables. As a result, while expert knowledge of the system is still needed for an ML model to be applied successfully, detailed information about the system is not, making these algorithms very promising to predict quantities of interest in complex systems such as those in focus here. This study aims to explore the ability of ML algorithms to predict flow boiling heat transfer coefficients in microchannels. 3. METHODOLOGY The framework for data regression proposed by Loyola-Fuentes et al. [5] is used here to analyse and predict the heat transfer coefficients reported in the Brunel Two-phase Flow Database. This extensive database comprises 27,126 data points for flow boiling in single and multi-microchannels of various materials, shapes, dimensions and lengths. The methodology used comprises six steps: 1) Data pre-processing, 2) Features and target variable selection, 3) Selection of training and testing data, 4) training and testing stages, 5) Hyperparameter tuning, and 6) Performance assessment [4]. 4. RESULTS First, a preliminary analysis has been carried out on the dataset to provide a statistical summary includingthe main features of the data and the variables ’ distribution. Figure 1 illustrates the raw data distribution of the variables that have been measured during the experiments. Figure 1: Raw data distribution of input variables and heat transfer coefficient. Filtering procedures have been applied to remove out-of-scope, i.e. single-phase, multichannel, and rectangular-shape

FAQs

Why are physics-based models challenging for flow boiling in microchannels?

Flow boiling involves multiple interacting microscale phenomena that are extremely difficult to capture mathematically with physics-based models.

What data and variables are used to train the machine learning models?

The approach uses the Brunel Two-phase Flow Database and extracts explanatory variables required for prediction, then applies regression models to estimate the heat transfer coefficient.

How is model performance evaluated and what limitation is observed?

The model is trained on 80% of the data and tested on 20%, with error metrics such as MAPE and residual analysis. The results show a tendency to underestimate HTC for values above 25,000 W/m²K, likely due to fewer experiments at that range.

Aug 6, 2026

Upload by Lucas Martin