Date of Graduation

12-2021

Document Type

Thesis

Degree Name

Master of Science in Mechanical Engineering (MSME)

Degree Level

Graduate

Department

Mechanical Engineering

Advisor

Han Hu

Committee Member

Darin W. Nutter

Second Committee Member

Paul Millett

Keywords

Computer Vision, Heat Transfer, Machine Learning, Pool Boiling

Abstract

The understanding of bubble dynamics during boiling is critical to the design of advanced heater surfaces to improve the boiling heat transfer. The stochastic bubble nucleation, growth, and coalescence processes have made it challenging to obtain mechanistic models that can predict boiling heat flux based on the bubble dynamics. Traditional boiling image analysis relies on the extraction of the dominant physical quantities from the images and is thus limited to the existing knowledge of these quantities. Recently, machine-learning-aided analysis has shown success in boiling crisis detection, heat flux prediction, real-time image analysis, etc., whereas most of the existing studies are focused on static boiling images, failing to capture the dynamic behaviors of the bubbles. To address this issue, in the present work, a convolutional long short-term memory (ConvLSTM) model is developed to enable quantitative prediction of heat flux based on sequences of boiling images, where the convolutional layers are used to extract the features of the boiling images and the LSTM layers to identify the temporal features of the sequences. A convolutional neural network (CNN) model that is based on the classification of static images is also developed as a reference. Both models are trained with images of HFE-7100 boiling on silicon micropillar arrays at different steady-state heat fluxes. The results show that both CNN and ConvLSTM models have led to accurate predictions of heat flux based on the boiling images. In particular, the ConvLSTM model is shown to yield higher accuracy for heat flux predictions of completely unseen data, indicating a higher level of generality. Another focus of the present work is the forecasting capability of data-driven models using boiling images under transient heat loads. A CNN regression model is coupled with a one-dimensional LSTM model to enable a quantitative forecast of heat flux during boiling. The model is trained using image sequences of water boiling on planar copper surfaces with power ramp-up and has demonstrated a reliable forecasting capability.

Share

COinS