Implementation of a data-driven model for mesh-induced error corrections in CFD simulations of stirred tanks

Using computational fluid dynamics (CFD) for reactor design is an established area in chemical engineering, and the emergence of machine learning (ML) approaches offers new possibilities for CFD. This work reports the integration of data-driven ML models with CFD to increase the efficiency of simulations of stirred tanks. By predicting and correcting mesh-induced errors in coarse-grid CFD simulations, it is demonstrated that ML models can significantly improve simulation results, reduce computational costs, and maintain high accuracy. This approach involves training a Random Forest surrogate model using high-fidelity and low-fidelity data generated from CFD and applying it to predict and correct coarse-mesh simulation inaccuracies. For the case study of single-phase flow in a stirred tank, the machine learning model demonstrated good performance in various scenarios, including interpolation and extrapolation of error predictions, highlighting the potential of combining ML with traditional CFD methods for flow field reconstruction. The study also explores the effect of selected physical features, providing the optimal feature combination for model training.