Deep Learning Forecasts a Strained Turbulent Flow Velocity Field in Temporal Lagrangian Framework: Comparison of LSTM and GRU

Abstract

The subject of this study presents an employed method in deep learning to create a model and predict the following period of turbulent flow velocity. The applied data in this study are extracted datasets from simulated turbulent flow in the laboratory with the Taylor microscale Reynolds numbers in the range of 90 < Rλ< 110. The flow has been seeded with tracer particles. The turbulent intensity of the flow is created and controlled by eight impellers placed in a turbulence facility. The flow deformation has been conducted via two circular flat plates moving toward each other in the center of the tank. The Lagrangian particle-tracking method has been applied to measure the flow features. The data have been processed to extract the flow properties. Since the dataset is sequential, it is used to train long short-term memory and gated recurrent unit model. The parallel computing machine DEEP-DAM module from Juelich supercomputer center has been applied to accelerate the model. The predicted output was assessed and validated by the rest of the data from the experiment for the following period. The results from this approach display accurate prediction outcomes that could be developed further for more extensive data documentation and used to assist in similar applications. The mean average error and R2 score range from 0.001–0.002 and 0.9839–0.9873, respectively, for both models with two distinct training data ratios. Using GPUs increases the LSTM performance speed more than applications with no GPUs.

Turbulent Flow Prediction-Simulation: Strained Flow with Initial Isotropic Condition Using a GRU Model Trained by an Experimental Lagrangian Framework, with Emphasis on Hyperparameter Optimization

“…It is reported that GRU is faster and produces similar prediction results as LSTM with fewer data [4,11,12,15]. In this study, 6,225,457 tracking points are available just from the 0.4 s long period of the experiment.…”

“…This architecture establishes two significant features: the reset gate captures short-term dependencies and the update gate models' long-term dependencies in sequences [28]. is the hidden state from the previous step, X (t) is the current input, h (t) is a new hidden state, y (t) is the output, r (t) is the reset gate, z (t) is the update gate, g (t) is the candidate hidden state, σ is the sigmoid function, and tanh is the hyperbolic tangent function [15].…”

“…The models are coded in Python with the TensorFlow library [30,31]. The GRU model is set up with 100 layers and one dense layer, and Adam is specified as an optimizer [15]. The dataset was normalized by the MinMaxScaler transformation [32], scaling the minimum and maximum values to be 0 and 1.…”

“…From the previous study of the author's research group, LSTM, GRU, and Transformer models have been applied with 2,862,119 tracking points, with two sequential variable inputs (temporal feature) composed of V X and V Y to predict the V X and V Y [15,16]. Table 1 shows that the GRU-h model of this study is 20% faster than the GRU model with a smaller dataset, and it is 14% and 15% faster than the LSTM and Transformer models, respectively.…”

“…Furthermore, most works include pre-processing steps to identify the dominant features, such as POD or DMD [4]. Recently, Hassanian et al [15] used LSTM and GRU models to predict a turbulent flow with only temporal features. Moreover, the Transformer model, as an up-to-date DL technique, displays successful capabilities to simulate and forecast emerging unknown patterns of turbulent flow [16].…”