mixing-ann

This repository includes all the code programs, scripts and files that were used for a project about the development of an Artificial Neural Network trained with data coming from Lethe simulations of mixers.

It is divided into two folders. The folder sim100k contains the files that served to generate the 100k CFD simulations of mixing systems. The second folder, called ai, contains the files that served to pre-process the data and to train the Artificial Neural Network.

💻 CFD simulations (in folder sim100k)

Setting the Python methods

In the MixerSim.py file, we define three important methods. The first one is used to create the 100k samples, the second to generate the mesh and the third to gather the torque and to calculate the power number $N_p$.

1. generate_data_folders

This method uses the Latin Hypercube sampling (LHS) method to generate 100k samples of mixing systems. It takes in arguments the range of the features of the mixing systems (the dimensionless geometrical features and the Reynolds number $Re$). The LHS method comes from the library pyDOE. The method creates 100k different folders in which every mixers will be simulated seperately.

Using the library Jinja2, we copy the mixer.geo template in the folder and we replace the arguments that generates the mesh by the right geometrical ratios. We apply the same principle to the mixer.prm template. We replace the kinematic viscosity parameter by the right value in order to respect the Reynolds number.

2. launch_gmsh

With the same library Jinja2, this methods specifies the minimum and maximum characteristic length of the tetrahedral element of the mesh. It then launches the software Gmsh to generate the mesh of the mixing vessel. This method is usefull to perform a mesh sensitivity analysis.

3. get_torque_and_write_data

When all the Lethe simulations are done, this method gathers all the torque resulting from the simulations and then calculates the power number $N_p$ of every 100k mixers. At the end of its execution, the method generates a file of type .txt that will serve as the database fed to the ANN.

Launch simulations on the cluster

In the sim100k folder, some files are used to launch the methods decribed above and the software Lethe strictly on the Digital Alliance of Canada clusters. Every step of the methodology includes two files: one .sh file and one .py file. The .sh file will serve as the sbatch script to queue a job on the cluster and the .py file will serve as a script that execute the different methods.

1. Launching the folders

The launch_folders.sh is the sbatch command that submits the launch_folders.py script as a job. It executes the generate_data_folders method to create the 100k folders containing 100k different mixing configurations.

2. Launching Lethe

The launch_lethe.sh is the sbatch command that submits the launch_lethe.py script as a job. It executes the launch_gmsh method to generate the mesh of the mixer using Gmsh. Then, it runs the Lethe executable gls_navier_stokes_3d build on the cluster.

3. Launching the gathering of the data

The launch_data.sh is the sbatch command that submits the launch_data.py as a job. It executes the get_torque_and_write_data to gather the features, the torque and ultimately the power number of every simulations in the final database.

🌐 ANN (in folder ai)

Setting the Python methods

In the MixerNN.py file, three important methods are implemented. The first one reads the database, the second normalizes the data and the third trains the ANN.

• read_mixerdata

This method reads the mixer_database_0-99999.txt file that contains the features (inputs) and the power number (output) of the 100k mixers simulations and stores them into variables.

• initial_setup

This method creates a tensor $X$ of dimension $n \times d$ where $n$ is the number of mixing samples and $d$ is the number of features, which is seven (six geometrical ratios and one Reynolds number). It also creates a second tensor $y$ of dimension $n \times 1$ containing the outputs $N_p$ of every mixing samples.

Then, we use the MinMax method from sklearn to scale the features and the outputs in the interval of 0 to 1 to avoid any bias during the training.

Finaly, we seperate the samples into two sets: training and testing. We use the train_test_split function from sklearn to perform the split.

• fit_model

Firstly, the method creates the architecture of the ANN using the Tensorflow and Keras libraries. It builds a deep network according to the arguments of the method that specifies its hyperparameters, such as the number of layers, the number of neurons, the batch size, the number of epochs and the activation function.

Secondly, the method compiles the ANN using the training set. In the method, we return the model and also the history of the ANN. This allows to catch the evolution of the loss function.

Grid search

The grid_search.py script performs a grid search with cross validation (GridSeachCV from sklearn) to find the best hyperparameters of the ANN.

Optimum training

We train the ANN with all the training set using the optimum.py script. It does the training using the hyperparameters determined by the grid search and returns the mean squared error (MSE), the mean absolute error (MAE) and the mean absolute percentage error (MAPE) of the ANN. Also, the final model is saved in the optimum_mixer_model folder.

Comparison

The first script (compare.py) generates power curves of $N_p$ versus $Re$ for a given geometry with different models. It compares the predictions of $N_p$ made by the ANN, the correlation of Hiraoka et al. and Lethe simulations. The data of the simulations are stored in Np-Re1_0-20.txt and Np-Re2_0-20.txt. The first set of simulations used a coarser mesh while the second used a finer mesh.

The second script (plot_data.py) predicts $N_p$ with the ANN and the correlation for multiple different geometrical configurations of the testing set. It utimately plots the comparison between the $N_p$ predictions and the real values coming from the database.