A deep learning framework for predicting and optimizing flow fields in reactive flows

Computational Fluid Dynamics (CFD) is widely used for solving and optimizing the flow fields of different systems and applications. However, running CFD simulations, especially for reactive flow systems, can be very time consuming and memory intensive, which limits e.g. design space exploration in the optimization tasks. In this work, a data-driven modeling methodology has been developed to predict 2D flow distributions in a chemical reactor. The primary objective was establishing correlations between global boundary conditions (including process and geometrical parameters), and the resulting CFD flow field distributions. A convolutional autoencoder was used to compress and reduce the data dimensions efficiently. Simultaneously, a multilayer perceptron served as the mapping mechanism that linked the global boundary conditions to the compressed data. The methodology developed in this work provides a very successful demonstration of its ability to map both geometric and process parameters to flow fields. The results showed a prediction accuracy of approximately 94%–97% for the CFD cases, indicating a very high prediction quality. Besides this, the prediction time was less than a second, which is significantly lower compared to the computational effort required for CFD simulations. To demonstrate the practical applicability of this approach, an interactive tool was developed to enable real-time visualization of predicted flow fields. This tool represents a foundational step towards applying digital twins and integrating such models into industrial practice.