This paper presents a novel methodology for optimizing the position of a propeller in an agitated tank using a Machine Learning (ML) model trained with Computational Fluid Dynamic (CFD) simulations. A hybrid model is used integrating multi-layer perceptrons (MLP) with convolutional layers (CNN) to enable rapid prediction of 3D velocity fields based on the spatial location of the propeller. This novel methodology allows evaluating advanced mixing metrics in milliseconds compared to traditional approaches based on slow CFD simulations or the prediction of reduced-order quantities. A 2D slicing strategy is used to enhance the training efficiency while preserving the 3D flow structures. This approach outperforms traditional 3D architectures and dimensionality reduction techniques The surrogate model is integrated in an optimizing tool to propose the best suitable position to maximize mixing efficiency. Best mixing scenario reduced the amount of dead volume by 89% compared with a neutral baseline design (from 62.9% to only 6.9%). Mixing performance was further investigated by means of a transient simulation involving virtual tracers. Two mixing metrics, namely Uniformity Index and second order central moments were employed, the results show that the optimized case required half the time to reach full homogeneity compared with the baseline. Finally, this new methodology proved to be more computationally efficient than a pure CFD-based optimization strategy, both when optimization candidates are evaluated sequentially or in batches to save computational time. • Optimization of mixing in agitated tank is done by means of Machine Learning (ML). • ML model predicts velocity field based on propeller location inside the tank. • Dead volume is decreased by 89% and homogeneity is reached 2 times faster in the optimal scenario.
Gómez et al. (Wed,) studied this question.