What is it about?
This study looks at a key thermal design problem for hypersonic vehicles: estimating the heat flux at the stagnation point, where heating is most severe. Normally this requires computational fluid dynamics (CFD), which is accurate but slow and costly. We tested whether a physics-informed neural network (PINN) could learn this relationship faster by combining simulation data with a known engineering heat-transfer correlation. The study used an axisymmetric blunt-nosed body with different nose radii, Mach numbers, and altitudes. Compared with a standard deep neural network, the PINN gave more accurate predictions and generalized better to new cases, while keeping the speed advantage of a surrogate model over repeated CFD runs.
Featured Image
Why is it important?
Thermal protection design for hypersonic vehicles depends on reliable heat-flux estimates, but full CFD can be too expensive to use repeatedly in early-stage design studies. The value of this work is that it shows how domain knowledge can be embedded directly into a learning model, improving both accuracy and robustness compared with a purely data-driven neural network. In the paper, the PINN outperformed the baseline DNN on the held-out test set and also on unseen internal and outside-domain cases, indicating better generalization. That makes this approach promising for faster screening, design exploration, and preliminary aerothermal assessment where many candidate conditions must be evaluated efficiently.
Perspectives
In this work, we wanted to explore whether machine learning models for aerothermal prediction could be made both efficient and physically meaningful. Rather than using a purely black-box model, we incorporated an established engineering correlation into the training process so that the network would be guided by domain knowledge. One of the most encouraging outcomes was that the physics-informed model remained more robust than a standard neural network even for unseen conditions, including cases outside the training range. We see this as a useful step toward practical surrogate models that can support faster thermal design studies in hypersonic applications.
Huseyin Avni Yasar
Middle East Technical University
Read the Original
This page is a summary of: Physics Informed Neural Networks for Enhanced Critical Heat Flux Prediction in Hypersonic Flows, July 2024, American Institute of Aeronautics and Astronautics (AIAA),
DOI: 10.2514/6.2024-4203.
You can read the full text:
Contributors
The following have contributed to this page
