Paper Submission
01. Experimental/Computational Fluid Dynamics
Improving Interpretability in Flow Field Mode Decomposition via Human-Guided CNNs
Mode decomposition is a technique used to extract significant features from flow fields, and recently, deep learning (DL) has been employed for this purpose. Conventional DL methods, such as spatio-temporal mode decomposition models using Convolutional Neural Networks (CNNs), can decompose time series of flow fields into nonlinear modes. However, the decomposed fields can vary with each training due to the stochastic nature of the learning process, leading to both expressive richness and interpretive difficulty.
To enhance interpretability, we propose a Human-Guided CNN Mode Decomposition Model in this study. This model constrains modes in the latent space, allowing decomposed flow fields to be derived from these constrained modes intentionally set by human input. The model architecture includes an encoder and a decoder. The encoder compresses time series of flow fields into a limited number of modes, trained to match supervised modes provided during training. The decoder then expands each mode into corresponding decomposed flow fields, and their summation approximates the original input flow.
The model is applied to a cylindrical flow at a Reynolds number, ReD=U∞D⁄ν=100, as an example of unsteady flows. We provide three sine waves with different periods as supervised modes in the latent space. These periods are set to match the shedding period of Karman vortices, as well as half and one-third of that period. The trained modes closely matched the supervised modes, and the reconstructed flow fields showed high accuracy. The decomposed fields reflected the characteristics of the latent space modes: the lowest frequency mode corresponded to Karman vortices, while the highest frequency mode related to steady structures. These findings validate the model’s effectiveness in producing interpretable decomposed flow fields by utilizing a constrained latent space determined by human guidance, thereby enhancing the interpretability of DL-based mode decomposition models in flow field analysis.
Download the file you uploaded
Author Information
Yosuke Shimoda
Mr.
Corresponding author, Presenting author
Naoya Fukushima
Prof.
Corresponding author