Prediction of particle trajectories in DNS with machine learning*

The transport of passive particles in turbulent flow can be studied with a combination of Direct Numerical Simulation (DNS) and Lagrangian scalar tracking (LST). While such methods are computationally expensive,¹ deep learning has become a promising approach for the prediction of tunrulent flow behavior results for  computational savings.² In the Lagrangian framework, spatial and temporal features are required to construct the particle movement. Long short-term memory (LSTM) and hybrid methods have been used for temporal evolution while training deep neural networks.³ Herein, we use LSTMin combination with artificial neural network (ANN) to generate the trajectories of passive particles. The data was obtained from DNS/LST computations for channel flow at Re_τ=300 and 100000 particle trajectores⁴ divided into 80% for training, 10% for validation and 10% for testing. The performance of the model was investigated at various hyperparameters and the mean square error was chosen as the loss function for fitting procedures. The prediction results were compared with the ground truth values based on the distribution of position and velocity over time. It was found that the LSTM exhibiteed high accuracy predictions at short times. 

 References

1. R. Hassanian, et al., Physics of Fluids 35, 075118 (2023)

2. Z. Xie, et al., Ind. Eng. Chem. Res.  61, 8551−8565 (2022)

3. G. Borelli et al., Int. J. Heat & Fluid Flow 96, 109010 (2022)

3. O. Pham et al., Phys Fluids 36(1), 015133 (2024)