Session: 11-03-01: Artificial Intelligence, Machine Learning and Data Science for Thermal Processes, Heat Transfer and Energy Systems

Paper Number: 145419

145419 - Machine Learning Assisted Models for Predicting and Optimizing Boiling Heat Transfer on Scalable Random Surfaces 

Boiling is a ubiquitous process in numerous applications, including nuclear power plants, water treatment, and thermal management. Predicting and enhancing boiling heat transfer performance is difficult due to the complex nature of the boiling process. Existing experimental correlations offer limited guidelines, while numerical simulations require extensive computational resources and several crucial assumptions. To overcome these challenges, recent studies have attempted to predict the boiling performance with the aid of the machine learning models, but offer limited capability of surface design optimization. Here, we develop supervised machine learning (ML) models based on convolutional neural networks (CNN) to predict boiling heat transfer performance using detailed surface information as the only input. We investigate a scalable sandblasting manufacturing technique for the enhancement of boiling heat transfer. We fabricated a large number of statistically distinct random surfaces by varying manufacturing parameters including particle size, inlet pressure, nozzle-to-sample distance, sweeping speed, and sweeping distance. Each surface was characterized by a 3D optical profilometry and classified using T-distributed stochastic neighbor embedding (t-SNE) method to visualize the effect of each manufacturing parameter. Five identical test setups were built and automated to obtain boiling curves of different samples in parallel. The pool boiling experiments were conducted with saturated water in an ambient laboratory condition to acquire sufficient ML training dataset. The supervised ML models were trained by feeding a variety of surface morphologies and boiling behaviors and validated to predict boiling performance solely based on surface morphology. For a given unobserved boiling surface, the CNN algorithms predicted both the boiling curves and the critical heat flux (CHF) values within a deviation of 10% from the experimental results, which outperformed classical boiling experimental correlations. We further developed an inverse design pipeline based on a generative adversarial network (GAN) to create the boiling surfaces for the target heat transfer performance. The GAN models generated the surface features with corresponding manufacturing parameters while the prediction model screens the candidates to identify the generated surfaces that will match the target boiling performance. Once the inverse design pipeline proposes the best candidates, we fabricated the boiling surfaces using the given manufacturing parameters and conducted the pool boiling experiments. The results showed that the boiling performance of the generated surfaces captured the trends of the different target boiling curves. This work not only provides optimal boiling surface design guidelines, but also will enable scalable surfaces to be more readily adopted and enhancing performance in various energy applications.

Presenting Author: Hyeongyun Cha University at Buffalo

Presenting Author Biography: Dr. Hyeongyun Cha is an assistant professor at the University at Buffalo (UB), The State University of New York. He received his doctoral degree in mechanical engineering from the University of Illinois Urbana-Champaign (UIUC). Before joining UB, he was a postdoctoral associate in the Department of Mechanical Engineering at the Massachusetts Institute of Technology (MIT).

Authors:

Hyeongyun Cha University at Buffalo
Yihao Zhang Massachusetts Institute of Technology
Yang Zhong Massachusetts Institute of Technology
Ziqi Lu Massachusetts Institute of Technology
Youngsup Song University of Florida
Minna Wyttenbach Massachusetts Institute of Technology
Amir White Massachusetts Institute of Technology
John Leonard Massachusetts Institute of Technology