Time-resolved turbulent velocity field reconstruction using a long short-term memory (LSTM)-based artificial intelligence framework

Abstract

This paper focuses on the time-resolved turbulent flow reconstruction from discrete point measurements and non-time-resolved (non-TR) particle image velocimetry (PIV) measurements using an artificial intelligence framework based on long short-term memory (LSTM). To this end, an LSTM-based proper orthogonal decomposition (POD) model is proposed to establish the relationship between velocity signals and time-varying POD coefficients obtained from non-TR-PIV measurements. An inverted flag flow at Re = 6200 was experimentally measured using TR-PIV at a sampling rate of 2000 Hz for the construction of training and testing datasets and for validation. Two different time-step configurations were employed to investigate the robustness and learning ability of the LSTM-based POD model: a single-time-step structure and a multi-time-step structure. The results demonstrate that the LSTM-based POD model has great potential for time-series reconstruction since it can successfully recover the temporal revolution of POD coefficients with remarkable accuracy, even in high-order POD modes. The time-resolved flow fields can be reconstructed well using coefficients obtained from the proposed model. In addition, a relative error reconstruction analysis was conducted to compare the performance of different time-step configurations further, and the results demonstrated that the POD model with multi-time-step structure provided better reconstruction of the flow fields.

Prediction of the dynamics of a backward-facing step flow using focused time-delay neural networks and particle image velocimetry data-sets

“…Srinivasan et al [47] compared deep feed-forward and recurrent LSTM networks for turbulent shear flows prediction. Recently Deng et al [48], used LSTM networks to reconstruct the POD coefficient time series using sub-sampled distributed velocity sensors in an inverted flag flow PIV experiment. A short review of applications of deep NN to fluid mechanics can be found in [49].…”

Hybrid deep neural network based prediction method for unsteady flows with moving boundary

“…The fully connected neural network was used to predict the time coefficients of POD modes. Similarly, Rahman et al (2019), Deng et al (2019) and Ahmed et al (2019) used LSTM to predict the time coefficients of POD modals, and Miyanawala et al (2019) used CNN to predict the time coefficients of POD modals. And combinations of the CNN and ConvLSTM also been proposed for dimensionality reduction and spatial-temporal modeling of the 3 / 22 flow dynamics by Mohan et al (2019), Hasegawa et al (2019) and Han et al (2019).…”

Data-driven order reduction and velocity field reconstruction using neural networks: The case of a turbulent boundary layer

“…It is generally necessary to use an order reduction to derive some information that could be handled either to model or control the targeted flow. Data-driven methods are nowadays becoming more and more efficient and reliable even for fluid mechanics research 9,17 . Among successful applications, one can cite statistical learning 16 or machine learning 14,35 algorithms.…”

“…In an inverted flag flow experiment, Deng et al 9 applied an ANN identification method to reconstruct time-resolved velocity fields from a handful of velocity sensors. It was also proven recently that ANNs can be used to predict the dynamics and reconstruct the time-resolved fields of an experimental Backward-Facing Step (BFS) flow 15 .…”