A Machine Learning Approach to Classify Kinematics and Vortex Wake Modes of Oscillating Foils

TL;DR

This paper presents a machine learning model combining CNN and LSTM to classify vortex wake patterns behind oscillating foils, achieving up to 90% accuracy, aiding optimal energy harvesting device placement.

Contribution

It introduces a novel CNN-LSTM classification approach for wake patterns in oscillating foils, improving accuracy over previous methods.

Findings

Achieved 80% accuracy with initial grouping.

Updated group boundaries increased accuracy to 90%.

Demonstrated effective classification of vortex wake structures.

Abstract

Machine learning techniques have received attention in fluid dynamics in terms of predicting, clustering and classifying complex flow physics. One application has been the classification or clustering of various wake structures that emanate from bluff bodies such as cylinders or flapping foils, creating a rich diversity of vortex formations specific to flow conditions, geometry, and/or kinematics of the body. When utilizing oscillating foils to harvest energy from tidal or river flows, it is critical to understand the intricate and nonlinear relationship between flapping kinematics and the downstream vortex wake structure for optimal siting and operation of arrays. This paper develops a classification model to obtain groups of kinematics that contain similar wake patterns within the energy harvesting regime. Data is obtained through simulations of 27 unique oscillating foil kinematics for a total of 13,650 samples of the wake vorticity field. Within these samples three groups are visually labeled based on the relative angle of attack. A machine learning approach combining a convolutional neural network (CNN) with long short-term memory (LSTM) units is utilized to automatically classify the wakes into the three groups. The average accuracy on five test data subsets is 80% when the three visually labeled groups are used for classification. After analyzing the test subset with lowest accuracy, an update on the group division boundaries is proposed. With this update, the algorithm achieves an average accuracy of 90%, demonstrating that the three groups are able to discern distinct wake structures within a range of energy harvesting kinematics.