Session: 1.2 - Machine Learning, Reduced Order Modeling in CFD and Design Optimization

Paper Number: 155524

155524 - Evaluating Machine Learning-Enhanced Sub-Grid Scale Stress Models With Invariance Embedding for Meso-Scale Hurricane Boundary Layer Flows 

Abstract: 

The complicated energy transfer events make it difficult to model turbulent flows in mesoscale atmospheric conditions, including hurricane boundary layers. Regarding energy backscatter, the crucial function that smaller scales play in larger scale energy flows, standard large-eddy simulation (LES) methods frequently overlook it. Traditional sub-grid-scale (SGS) stress models misrepresent hurricane boundary layer flow by focusing on energy dissipation and neglecting backscatter dynamics. 

This study uses SGS stress models improved with machine learning to address this gap and incorporate geometrical and physical invariances. We assess these methods' capacity to reproduce energy dynamics in LES, relying on MATLAB's Bayesian regularization-based optimizer ('trainbr') and moving to Python's ADAM optimizer. The feed-forward neural networks (FNNs) are trained to anticipate energy cascades and backscatter with different degrees of embedded invariances using high-resolution LES data of hurricane-like vortices. Model generalizability and forecast accuracy are enhanced by these invariances derived from flow physics.

To conduct this analysis, we will compare several neural network configurations with distinct GPU setups and optimization strategies. Initially, we will use MATLAB's Bayesian regularization-based optimizer ('trainbr') to evaluate the neural networks. Following this, we will transition the implementation to Python to leverage the flexibility and scalability of its deep learning frameworks. Specifically, we will utilize the ADAM optimizer in Python and compare its computational efficiency and accuracy to MATLAB's trainbr optimizer. Our evaluation will focus on key performance metrics, including the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE). We will analyze the computational trade-offs between MATLAB and Python implementations, particularly in the context of large-scale simulations requiring significant data processing. By transitioning from MATLAB to Python, this study highlights the adaptability and scalability of Python for machine learning applications in atmospheric simulations on GPU clusters. Python's robust libraries and compatibility with high-performance computing frameworks provide an ideal platform for integrating machine learning into operational workflows. Through this transition, we aim to demonstrate the potential of Python-based implementations to facilitate more efficient and scalable turbulence modeling.

This study emphasizes the shift away from frameworks based on MATLAB and toward implementations written in Python. In addition to this, it places an emphasis on the adaptability and scalability of Python implementations for applications that use deep learning and considerable amounts of data. These findings highlight the potential of machine learning-enhanced SGS models to improve the realistic depiction of LES, hence laying the groundwork for a posteriori testing to be conducted in genuine atmospheric simulations. Ultimately, this research aims to advance the development of next-generation SGS models capable of addressing the complexities of mesoscale turbulence. These findings will provide valuable insights into the application of machine learning in LES, paving the way for more accurate and efficient simulations of hurricane boundary layers and other extreme weather phenomena.

Presenting Author: Md Badrul Hasan University of Maryland, Baltimore County

Presenting Author Biography: Mr. Md Badrul Hasan joined Dr. Meilin Yu's lab in 2019 after completing his undergraduate studies at the Bangladesh University of Engineering and Technology (BUET). He has worked on analyzing the effects of numerical dissipation from different numerical weather prediction models on hurricane intensification simulation for his Master's dissertation. His current research interests are in high-fidelity numerical simulation, machine/deep learning, and their applications to environmental fluid dynamics.