Session: K8-01: FUNDAMENTALS OF MACHINE LEARNING IN HEAT TRANSFER

Paper Number: 131218

131218 - A Convolution Neural Network Design for Combined Image and Sensor Data Analysis to Determine Droplet Vaporization Regime and Heat Transfer Performance 

Abstract: 

Combining high speed video cameras and optimal measurement techniques with digital sensors controlled by a data acquisition system can yield a powerful combination of experimental tools to explore boiling process thermophysics and heat transfer mechanisms.  More specifically, for boiling systems, imaging can provide information on the two-phase morphology of the system, both in the near wall region where nucleation, bubble growth and release affect heat transfer, and farther from the surface where two-phase flow and transport behavior can also affect system heat transfer performance.  Imaging can provide qualitative and quantitative information that complements data provided by temperature, pressure and velocity sensors.  This type of combined use of imaging and digital sensor instrumentation can be further enhanced by using machine learning tools to analyze the flow of information from imaging and digital sources.  It is, however, challenging to optimally combine these information flows in machine learning analysis.  Digital data is easily collected and analyzed in sequential neural networks, whereas convolution neural networks are commonly use to extract features and information from images.  This paper summarizes results of an exploration of machine learning strategies to optimally combine and analyze boiling process image and digital sensor information from experiments. Our early work on this topic used a collection of simulated images, created with features typical of boiling systems.  This allowed us to precisely include features in a controlled way so we could assess how those features were detected and merged with digital information by different machine learning process designs.  We specifically sought a machine learning process design for use in analyzing vaporization of deposited water droplets on superheated surfaces that may have varying degrees of nucleate boiling effects.  We needed a machine learning strategy that could combine image morphology information and digital measurements of pressure, superheat temperature difference, droplet initial size, surface contact angle, liquid wicking speed, and droplet evaporation time.  The system objective was to extract from this data the regime of vaporization (conduction driven only, conduction plus nucleate boiling, or explosive boiling), the liquid morphology (spread flattened layer, curved variable thickness droplet, droplet with nucleating and growing bubbles, shattered segmented liquid), and the capability to predict the vaporization regime and the wall superheat as a function of two-phase morphology and operating system parameters.  Through experimentation we found that a hybrid parallel-series convolution/neuron neural network design worked very effectively.  The network developed has two identifiable paths taken by the input images. Both paths, for classification and regression, have similar structure. They go from a double stacked convolution 2d layer to a ReLU activation then max pooling. Then they undergo flattening. From this point, the bubble classification and morphology classification take two distinct paths in interacting with a dense layer before concatenating with the input pressure and making predictions. This network design could be trained to predict regimes and evaporation time to better than 95% accuracy.  In subsequent work, we performed droplet vaporization experiments on superheated test surfaces, imaging them with high-speed video and recording operating conditions as digital measurements.  We used the best-performing hybrid design we developed using the simulated data and images to analyze the combined image and data flow from our real droplet experiments, and demonstrated that this hybrid network design successfully analyzed the data and could predict the vaporization regime and corresponding wall superheat to an accuracy of better than 92%.  The hybrid network developed in this research appears to be promising strategy for analyzing experimental data for physical systems that are best investigated experimentally with combined use of imaging and digital sensor instrumentation.

Presenting Author: Ursan Tchouteng Njike University of California at Berkeley

Presenting Author Biography: NSF Fellowship recipient Ursan Tchouteng Njike is a Graduate Student Researcher in the Energy and Multiphase Transport Laboratory at UCBerkeley.

Authors:

Ursan Tchouteng Njike University of California at Berkeley
Anisa Silva University of California at Berkeley
Van Carey University of California at Berkeley