Session: 08-05-01 Deep Learning for Predictive Modeling in Marine and Offshore Systems

Submission Number: 157344

Reduced-Order Modeling of Flow Around Cylinders Using Machine Learning 

This research focuses on the development and application of advanced Reduced-Order Modeling (ROM) methods to address fluid dynamics problems in marine and ocean engineering. The reduced-order models (ROMs) generated from these methods will provide efficient and accurate simulation of flow fields. However, while traditional Computational Fluid Dynamics (CFD) models offer high accuracy, the complexity of flow fields results in significant computational resource demands. ROM methods, by simplifying the problem while maintaining sufficient accuracy, provide an effective solution that can significantly reduce computational costs. All algorithms in this study are developed and tested based on two-dimensional cylinder flow fields. In this study, proper orthogonal decomposition (POD), neural networks (NNs), autoencoders (AEs), and radial basis function neural networks (RBFNN) will be employed to perform dimensionality reduction and feature extraction for flow fields data under different conditions. Numerical tests have been conducted in flow fields to compare three ROM methods: POD-ROM method, AE-ROM method, and POD-AE-ROM method. The results from two-dimensional cylinder flow data show that: (1) the POD-ROM method is computationally efficient but suffers from lower accuracy due to the loss of higher-order modes; (2) the AE-ROM method can produce results which have good agreement with the actual flow field during reconstruction; however, it loses physical information during the dimensionality reduction process, making it unable to fully preserve the physical characteristics in the reduced-dimensional space. (3) the POD-AE-ROM method combines the strengths of both POD-ROM method and AE-ROM method, maintaining predictive accuracy while also preserving physical characteristics with the highest level, therefore demonstrating superior performance. In addition, Long Short-Term Memory (LSTM) neural networks are employed to simulate and predict the time series of flow fields reduced by POD-ROM, AE-ROM, and POD-AE-ROM methods. The results from two-dimensional cylinder flow data show that: (1) in the POD-ROM flow field, the prediction accuracy of the neural network modeling the time series of POD coefficients decreases as the simulation progresses; (2) in the AE-ROM flow field, the prediction accuracy remains high throughout the entire simulation period but suffers from the loss of physical information during dimensionality reduction, preventing the reduced-dimensional space from fully retaining the flow field's physical characteristics; (3) in the POD-AE-ROM flow field, the introduction of AE reduces the dependency on higher-order POD modes, enabling high prediction accuracy across the entire simulation period. At the same time, the retention of primary POD modes ensures that the physical characteristics of the flow field are preserved to the maximum extent, making the POD-AE-ROM method superior in terms of overall performance.

Presenting Author: Guangyao Wang University of Macau

Presenting Author Biography: Dr. Guangyao Wang is currently an Assistant Professor in the Department of Ocean Science and Technology at the University of Macau. His research mainly focuses on the study of ocean dynamics modeling and renewable energy harvesting.