ANN to correct coarse-mesh CFD errors in cell culture bioreactors

Published: 30 November 2023| Version 1 | DOI: 10.17632/82k4wv6hhf.1
Fernando Cantarero,


This data was generated from CFD models of a spinner flask developed in COMSOL with two different impeller geometries and varying rotational speeds. The 1st derivative, 2nd derivative, and cell Reynolds data are used as inputs to train the ANN model. The fine and coarse velocity data are used to calculate the mesh error and used as training output. Then, the trained ANN is used on a set of testing data to predict mesh errors and correct the test coarse data. Results are saved as a correlation plot between corrected and coarse data against fine data and also as a table containing correlation coefficients, MSE, RMSE, and normalized RMSE. If the corrected results have lower RMSE than the coarse data, the ANN model has worked properly in correcting the coarse data.


Steps to reproduce

The Deep Learning Toolbox is required in Matlab to be able to run the code. The datafiles for each geometry and the viscosity case are found within their corresponding folders. Step 1: Ensure that the desired scenario's coarse and fine data files are copied into the "data" folder (e.g., the coarse and fine data for the original geometry.) Step 2: Once files to use have been copied into the "data" folder, ensure the matlab code is using the corresponding variable names (viscosity or rpm). Step 3: Decide on what training and testing datasets to use and run code. Step 4: Results will be saved in newly created folders named "Prediction on... [viscosity or rpm tested]". Save these files in a different location before running the code again, or they will be overwritten.


University of Tennessee Knoxville


Artificial Neural Network, Machine Learning, Computational Fluid Dynamics


AI Tennessee Initiative at the University of Tennessee