TL;DR
Created as part of my internship at Institute for Plasma Research, Gandhinagar, Ahmedabad.
We have implemented Deep Neural Network architectures as proposed by Srinivasan et al. using TensorFlow and made an attempt to reproduce the results as described in the study.
Detailed Explanation
The project aims to establish a workflow for modeling the turbulent flow of fluids using DNNs. Specifically, Recurrent Neural Network (RNN) will be employed for the above stated task.
The first task is to compare the performance of RNNs with Long Short Term Memory (LSTM) to quantify the superior capabilities of LSTMs in handling time series data, a crucial aspect of the project.
Subsequently, the project will involve generating parameterized time-series data featuring Moehlis coefficients. Validation of this data is essential to ensure the absence of laminarization over time.
The project’s final phase will encompass the training of the LSTM model using the generated dataset. The predictions made by this model will be validated against established physical metrics. This approach ensures that the model’s outputs align with Moehlis model which is a reduced dimensional solution of the Navier-Stokes equation governing the underlying fluid turbulence.
The datasets generated consists of the values of Moehlis coefficients at different points in time.
These are derived by solving a set of nine ordinary differential equations. The amplitudes are obtained for 4001-time points, resulting in a time series data. The required number of datasets can be generated, depending on the input provided.
Read the TODO to see the current task list.
The python script Turbulent_Shear_Flow_Prediction.py is still under development. Until then please refer the Notebook Version.
This project is licensed under the Apache License - see the LICENSE file for details.