Session: 17-01-01 Research Posters

Paper Number: 76510

Start Time: Thursday, 02:25 PM

76510 - Machine Learning Approach Towards Heat Transfer Correlations in Rough Cooling Channels 

We have explored machine learning approach as a heat transfer correlation. Convective heat transfer often manifests in complex physics such that empirical correlations i.e., mathematical forms that correlate convective heat transfer with surface geometry and flow conditions, are popular means to solve the convection problems. However, channels with rough surfaces, such as wavy channels, corrugated channels and channels integrated with internal ribs are associated with high order non-linear convective heat transfer which is challenging to represent using simple mathematical forms with high accuracy. Machine learning, ML, can be used be to analyze a large amount of data obtained from experiments, field measurements and numerical simulations. ML based data analysis not only improves the throughput and accuracy of flow interpretation, but it can also predict flow properties based on qualitative data or data from past occasions. Our study focuses on a cooling channel integrated with variable rib roughness. Roughness in the channel is controlled by varying the ribs height and spacing between them. A parametric analysis for 243 different rib array geometries is performed using finite volume model, FVM, by using adequate combination of rib height and pitch to generate training data set.  The ML model is first trained by FVM data using random forest regressor and predicts the local convective heat transfer coefficient of the cooling channel. The hyper-parameters are tuned to enhance the accuracy. Then, the interpolation capability of the random forest regression is tested using various new channel geometries. Despite the high complexity and nonlinearity of convective heat transfer coefficient along the rough channel, a single RF model predicted closely to FVM calculations with coefficient of determination, R² > 0.860 for the training data set and R² > 0.966 for the testing dataset, which has not been accomplished by any single mathematical form correlation. For the new validation dataset, i.e., designs unseen during ML training, the random forest algorithm was also able to predict convective heat transfer coefficient, similar to FVM with R² ~ 0.99 based on its interpolation mechanism.  ML model was proven to be versatile, as refining the resolution or expanding the range of parameters became much simpler task relative to the traditional correlation development.  For more precise prediction of refined regions, ML model was retrained by integrating new data into an original training data, which accurately predicted convective heat transfer coefficient of new channel designs. The demonstrated framework is widely adopted in various fields with high reliability; thus, it can be extended to other heat transfer problems involving multiple variables and complex phenomena.

Presenting Author: Faizan Ejaz ASU

Authors:

Faizan Ejaz ASU
Leslie Hwang N/Aa
Beomjin Kwon Arizona State University