Session: 17-01-01: Research Posters

Paper Number: 150375

150375 - Spray Atomization Characterization With the Aid of Machine Learning 

Atomization, the process of breaking down a liquid stream into small droplets to form a spray, is vital in various engineering applications such as spray cooling, painting, plasma spraying, and metal powder manufacturing. In aviation, atomization is crucial for spray combustion in engines, involving fuel injection, droplet formation, evaporation, and mixing with air before combustion. The quality of atomization, assessed by droplet size, shape, and dispersion, is essential for the efficiency of these processes. However, the current state of the art in theoretical and computational modelling of sprays does not accurately capture the shape of the droplets within the sprays. Recent research by our research group and others has emphasized how accurate modelling of the shape of droplets in sprays is important to predicting droplet (and therefore fuel) position and mixing in the turbulent flow.

This project investigates the atomization of a liquid jet in cross-flow using an in-house interface-capturing numerical simulation code. The research aims to enhance the characterization of atomization quality through the innovative use of machine learning methods. Volume of Fluid (VOF) simulations are conducted to extract detailed droplet information in cross-flow configurations. These simulations provide a comprehensive dataset, including droplet sizes, shapes, velocities, and spatial distributions. Techniques for identifying and analyzing the droplet interfaces and surface areas are employed to gain deeper insights into the atomization process.

This initial study focuses on extracting the sizes and shapes of droplets from spray and classifying shapes into a small number of classes. The results of this work will be used to develop novel models of droplets which reflect the true (non-spherical) shape of droplets found in sprays.

An algorithm that automatically detects and identifies droplets will be used to divide the VOF field into droplets. The PLIC (Piecewise Linear Interface Calculation) technique is used for interface reconstruction to accurately delineate droplet interfaces and compute geometrical features such as surface areas and volumes which are used as inputs to the machine learning algorithm. This method provides a reliable baseline for our analyses, ensuring that the droplet characteristics are captured with high fidelity. Enhanced algorithms for interface reconstruction are tested for improved precision in calculating geometrical properties. Several machine learning strategies are investigated for their ability to analyze and classify droplet shapes. Off-the-shelf methods from Scikit Learn, an open-source machine learning library in Python, are investigated including Support Vector Machines (SVM), Random Forests,  Gradient Boosting Machines (GBMs), and Principal Component Analysis (PCA).

This research represents a significant step towards optimizing atomization processes through the integration of CFD and machine learning, promising improved performance and efficiency in spray-dependent technologies. The project leverages advanced computational techniques and machine learning to improve the characterization and prediction of atomization quality. By focusing on accurate interface reconstruction and innovative data analysis methods, we aim to enhance the efficiency and effectiveness of various spray applications, ultimately contributing to advancements in jet fuel studies.

Presenting Author: Srinivasa Pavan Kancharla University of Washington

Presenting Author Biography: I'm a second-year master's student in the Department of Mechanical Engineering at the University of Washington. With my continuing interest in the area of Data Science and Machine Learning, I'm pursuing the data science offered by the department. I'm happily conducting research in spray atomization with the Aid of Machine learning which is an intersection of both my interest areas and looking to use much deeper neural networks going ahead. Apart from this, I'm a sportsperson and I hike as well.

Authors:

Srinivasa Pavan Kancharla University of Washington
Yushu Lin University of Washington
John Palmore Jr University of Washington