Dynamics-disentangled deep learning model for multi-cycle prediction of unsteady flow field

Abstract

The prediction of an unsteady flow field inherently involving high-dimensional dynamics is challenging. The multi-cycle prediction is especially difficult due to the inevitably accumulated errors over time. A novel deep learning model is proposed in this paper to disentangle the high-dimensional dynamics into two separate attributes that, respectively, represent spatial and temporal dynamics. A continuous mapping of temporal dynamics is subsequently constructed, which alleviates the error accumulation and, thus, contributes to the long-term prediction of the unsteady flow field. The dynamics-disentangled deep learning model (D3LM) processes sequential image data of the unsteady flow field and is constituted by three sub-networks, an encoder introducing a stochastic latent variable to explicitly model the low-order temporal dynamics (called varying attribute herein) and extracting multi-level representations of spatial dynamics (called consistent attribute herein), a decoder integrating the disentangled attributes and generating a future flow field, and a discriminator improving the quality of the predicted flow field. The proposed model is evaluated by two simulated datasets of unsteady flows around a circular cylinder at divergent Reynolds numbers. Benefiting from modeling the continuous distribution of temporal dynamics with the stochastic latent variable, the proposal can give multi-cycle future predictions with high accuracy both spatially and temporally on the two datasets with a small amount of training data. Our work demonstrates the potential practicability of deep learning techniques for modeling the long-term nonlinear laws of unsteady flow.