Analysis of chaotic patterns in Kelvin-Helmholtz instability by deep learning

• Main Author: Huang, Xinyuan
• Other Authors: Chow Wai Tuck
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Mechanical engineering::Fluid mechanics
• Online Access: https://hdl.handle.net/10356/168109
• Institution: Nanyang Technological University

Kelvin-Helmholtz (KH) instability is a complex fluid instability that occurs at the interface between two fluids moving at different velocities, such as the atmosphere and ocean, or in the wake of an airplane or ship. This instability produces characteristic swirling patterns that can be challenging to analyze due to several reasons: The Kelvin-Helmholtz instability exhibits nonlinear behavior, as the interaction of vortices can result in the formation of smaller vortices, ultimately leading to the disruption of the fluid interface. This phenomenon introduces complexities in accurately capturing the intricate dynamics of the instability. And a significant challenge in studying the Kelvin-Helmholtz instability lies in its multi-scale nature. The instability can occur at various scales simultaneously, posing difficulties in resolving all scales accurately during numerical simulations. This characteristic demands advanced computational techniques to effectively capture and analyze the multi-scale dynamics associated with the Kelvin-Helmholtz instability. Furthermore, the computational challenges associated with investigating the Kelvin-Helmholtz instability further complicate the study. Numerical simulations required to analyze this instability are computationally intensive, demanding substantial computational resources. The high computational costs make such simulations arduous and expensive to conduct, necessitating efficient algorithms and robust computing infrastructure to facilitate accurate and cost-effective investigations of the Kelvin-Helmholtz instability. Therefore, to analyze the KH instability accurately, a combination of computational fluid dynamics and deep learning methods are suggested to accurately simulate the results of KH instability under specific configuration and build a deep learning model for KH image recognition. In some cases, only recognition of a phenomena is required, the method of image recognition can be thousands of times faster than simulation algorithms. Computational Fluid Dynamics (CFD) is a multidisciplinary field that integrates modern fluid mechanics, numerical mathematics, and computer science, which is a dynamic discipline that employs computers as a tool to solve a wide range of practical problems. ANSYS FLUENT simulation is applied to acquire the data when initial velocity 1 = 0.25 / and 2 = 0.1 / , at real time slots of t = 16 s, 24 s, 32 s and 48 s and when initial velocity 1 = 0.25 / and 2 = 0.1 / , corresponding to the time slots at 8s, 16 s, 24 s of real time. At the same time, the data in each transient state under dynamic pressure, volume fraction and tangential velocity conditions are obtained for later deep learning test. The outputs indicate that the impact of the initial flow velocity on the Kelvin-Helmholtz instability problem is significant as the initial flow velocity is increased, the Kelvin-Helmholtz instability formation is accelerated. TensorFlow is an open-source software that developed for machine learning. Its flexible platform provides various features for building pre-trained machine learning models, including deep neural networks. In the realm of image recognition and classification, TensorFlow offers a variety of pre-trained models, such as Inception, ResNet, and MobileNet, that can identify objects and categorizing them into distinct groups. These models can be fine-tuned with specific image datasets to improve their accuracy and performance. The present study involves the utilization of a TensorFlow model for the purpose of identifying and classifying the Kelvin-Helmholtz problem. This model has demonstrated an exceptional capacity to maintain high levels of accuracy during testing samples, with a test reliability that reaches 100%. However, in the case of simulation data related to the Kelvin-Helmholtz problem, the reliability of the results is somewhat lower than that observed in the case of testing image instability. Despite this, the overall test accuracy is more than 70%, which is consistent with recognized standards for image recognition. Moreover, the simulation results obtained for the actual Kelvin-Helmholtz instability problem also demonstrate relatively high levels of accuracy during testing.