Machine Learning-Based Physics Inference from High-Fidelity Solutions: Vortex Core Data Extraction

Abras, Jennifer (HPCMP CREATE)

Nathan Hariharan


The field of machine learning is broad, covering many different areas and applications. The identification of aerodynamic flow features is one of many possible categories. The ability to train a machine learning model to classify aerodynamic flow features opens up a whole field of potential applications. Potential applications include automated data extraction, onboard solution assessment, and volume solution physics comparisons. The focus of the current effort is the demonstration of an automated machine learning-based vortex core data extraction tool. Vortical structures dominate the aerodynamics of the flow field beneath a hovering rotor blade. Years of archived research have been mined for these data to create a Convolutional Neural Network (CNN) model that effectively identifies vortices within computed CFD simulation results. An additional unsupervised learning tool takes in the information from the CNN model and extracts the vortex core properties. The combined tool automates what usually is a tedious manual process.

The discussion will cover the independent machine learning components of the utility and the complete process that utilizes these components. The independent machine learning components cover both vortex classification and vortex center localization CNNs and unsupervised learning techniques used for vortex assignment tasks. The methodology that wraps these independent machine learning techniques into a single utility that extracts core properties will be discussed before demonstrating applications of the utility to accurate CFD simulation results. The utility has presently been demonstrated on a wide variety of hovering rotor wake investigations, such as vortical-wake breakdown and impact of blade tip shapes on the wake structure.

The benefits of the automated utility over the manual process are demonstrated. The results presented show how machine learning may be applied overall to improve the post-processing workflows of the computational community.


Material presented in this abstract is a product of the CREATE (Computational Research and Engineering for Acquisition Tools and Environments) element of the U.S. Department of Defense HPC Modernization Program. In addition, the authors would like to acknowledge the support of the supercomputing resources provided by the HPCMP, in particular the Army Engineer Research and Development Center (ERDC) and the Air Force Research Laboratory (AFRL).