Date of Award

Fall 12-2022

Access Type

Thesis - Open Access

Degree Name

Master of Science in Aeronautical Engineering


Aerospace Engineering

Committee Chair

Mark Ricklick

Committee Co-Chair

Sandra K.S. Boetcher

Committee Advisor

Prashant Shekhar

First Committee Member

Habib Eslami


Supercritical carbon dioxide as a working fluid in a closed Brayton cycle is proving to be more efficient than a conventional steam-based Rankine engine. Understanding the heat transfer properties of supercritical fluids is important for the design of a working engine cycle. The thermophysical properties of supercritical fluids tend to vary non-linearly near the pseudo-critical region. Traditionally, empirical correlations are used to calculate the heat transfer coefficient. It has been shown in the literature and within our own studies that these correlations provide inaccurate predictions near the pseudo-critical line, where heat transfer may be deteriorated or enhanced, resulting from strong buoyancy and acceleration effects, and strong variations in fluid properties. The current study successfully uses machine learning techniques to capture these non-linearities and complex physics, providing an accurate tool for the design of heat transfer devices. The dataset is generated using highly validated computational fluid dynamics analysis. The bulk temperature and wall temperature data were obtained for a range of heat flux (q = [6, 12, 24, 36, 48]) and mass flux (G = [200, 400, 600, 800, 1000]) conditions. An artificial neural network base model was trained, validated, and tested using the CFD data. The test case was strategically selected such that the artificial neural network model trained on the high heat flux and mass flux (extreme) cases. Using the base model, hyperparameter tuning was performed, bringing down the prediction error on the test case by 94%. The final model predicted on the test set with an error less than 1%. This approach is computationally cost-efficient compared to the traditional correlation-based approach as it took only a few minutes for the model to train and predict. Lastly, this study published an artificial neural network tool that can be used to predict the wall temperature. Establishing a machine learning model capable of accurately predicting the wall temperature will aid in the design and development of future power generation cycles.