Affiliation: | 1. Department of Chemical Engineering, Polytechnique Montréal, Research Unit for Industrial Flows Processes (URPEI), Montréal, Québec, Canada Department of Chemical Engineering, Polytechnique Montréal, Engineering Process Intensification and Catalysis (EPIC), Montréal, Québec, Canada Contribution: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft, Writing - review & editing;2. Department of Chemical Engineering, Polytechnique Montréal, Research Unit for Industrial Flows Processes (URPEI), Montréal, Québec, Canada;3. Department of Chemical Engineering, Polytechnique Montréal, Engineering Process Intensification and Catalysis (EPIC), Montréal, Québec, Canada Contribution: Supervision, Writing - review & editing |
Abstract: | The power consumption of the agitator is a critical variable to consider in the design of a mixing system. It is generally evaluated through a dimensionless number known as the power number . Multiple empirical equations exist to calculate the power number based on the Reynolds number and dimensionless geometrical variables that characterize the tank, the impeller, and the height of the fluid. However, correlations perform poorly outside of the conditions in which they were established. We create a rich database of 100 k computational fluid dynamics (CFD) simulations. We simulate paddle and pitched blade turbines in unbaffled tanks from 1 to 100 and use an artificial neural network (ANN) to create a robust and accurate predictor of the power number. We perform a mesh sensitivity analysis to verify the precision of the values given by the CFD simulations. To sample the 100 k mixers by their geometrical and physical properties, we use the Latin hypercube sampling (LHS) method. We then normalize the data with a MinMax transformation to put all features in the same scale and thus avoid bias during the ANN's training. Using a grid search cross-validation, we find the best architecture of the ANN that prevents overfitting and underfitting. Finally, we quantify the performance of the ANN by extracting 30% of the database, predicting the using the ANN, and evaluating the mean absolute percentage error. The mean absolute error in the ANN prediction is 0.5%, and its accuracy surpasses correlations even for untrained geometries. |