Convolutional Neural Networks for Surrogate Modelling of Tube Bundle Heat Exchangers in Electrified Aircraft Propulsion

Kurzfassung

The aviation industry is undergoing a transformation toward sustainable propulsion technologies, with electrified aircraft propulsion powered by proton exchange membrane fuel cells emerging as a promising solution. These systems offer substantial reductions in greenhouse gas emissions and noise but face major challenges in thermal management. With typical fuel cell efficiencies of around 50%, a significant portion of input energy is released as waste heat. Ensuring stable operating conditions therefore requires compact and efficient thermal management systems (TMS), within which heat exchangers (HEXs) play a central role. The design of HEXs for electrified propulsion demands a delicate balance between high thermal performance, low pressure drop, reduced weight, and manufacturability. Traditionally, HEX analysis and optimization have relied on high-fidelity computational fluid dynamics (CFD) simulations, which accurately capture flow and heat transfer phenomena but incur prohibitive computational costs. This limitation restricts their use for parametric sweeps, sensitivity analyses, or large-scale design optimization. To overcome these barriers, this thesis develops a datadriven surrogate modeling framework that leverages deep learning to reproduce CFD-level accuracy at a fraction of the cost. Specifically, a geometry-adaptive convolutional neural network (CNN) based on U-Net architecture is introduced to predict the flow and thermal fields in staggered tube-bundle heat exchangers. A dataset of 400 two-dimensional Reynolds-Averaged Navier–Stokes (RANS) CFD simulations was generated using the SU2 solver across systematically varied geometric parameters, including tube diameter, longitudinal spacing, and transverse spacing. Each simulation provided velocity components, pressure, temperature, and turbulent viscosity fields, which were resampled onto structured grids and paired with geometry encodings (binary masks and signed distance functions). Five independent U-Net models were trained on these inputs, with specialized loss functions designed to capture both local flow features and global field consistency. The surrogate models achieved high predictive accuracy, with mean relative errors below 2% for velocity and temperature, and under 4% for pressure and turbulent viscosity fields. Importantly, the CNN-based framework reduced computational cost by approximately three orders of magnitude compared to CFD, enabling rapid evaluation of new HEX designs in under one second. Sensitivity analyses further confirmed that increasing dataset size improved robustness and minimized prediction variability across different geometries. In conclusion, this work demonstrates the effectiveness of deep learning surrogates for accelerating HEX design in electrified aircraft propulsion. By uniting the physical fidelity of CFD with the efficiency of machine learning, the proposed framework provides a scalable tool for design-space exploration and optimization, paving the way for reliable and lightweight thermal management solutions in next-generation sustainable aviation.