Physics-Guided Convolutional Neural Network for Flow Prediction in Heat Exchangers in Electrified Aircraft

Abstract

Electric aircraft propulsion systems powered by fuel cells are sensitive to operating temperature. With a roughly 50% efficiency of fuel cells, the requirement of heat rejection is a major challenge in the Thermal Management System (TMS). One of the most essential components of the TMS is the heat exchanger (HEX). To simulate the convective heat transfer inside the HEX, high-fidelity Computational Fluid Dynamics (CFD) simulations are commonly carried out. However, performing such simulations for sensitivity analysis and geometry optimization is computationally expensive. To overcome this, a deep-learning-based approach for surrogate modeling is pursued. A geometry-adaptive Convolutional Neural Network (CNN) surrogate model was constructed in the current study to predict the RANS mean flow fields of 2D HEX tube bundles with arbitrary cross-sections. Two CNN models were trained on the RANS simulations of 180 geometry samples to predict the velocity, pressure and temperature distribution. Primarily, a physics-informed CNN model was trained on a combination of the data-driven loss and the RANS governing equations. The spatial gradients of flow variables in this loss function were computed using a sixth-order central difference scheme on the uniform grid. Notably, the Dirichlet boundary conditions were strongly encoded in the CNN architecture and were thus inherently satisfied in the predictions. Additionally, a purely data-driven model was trained to compare the performance with the physics-informed approach. Both the CNN surrogate models were compared on qualitative and quantitative grounds. Overall, the RANS-guided model outperformed the simple data-driven model, thus reinforcing the advantages of this approach. On the unseen test geometries, the RANS-guided model predicted the distinct flow features in velocity, pressure and temperature with 97% accuracy and satisfied the governing equations as well as the boundary conditions. This approach not only alleviates overfitting and the need for large training datasets but also provides reliability. The RANS-guided surrogate model was developed to serve as a useful tool in the design exploration and optimization of HEX geometry in electrified propulsion systems.