Abstract
Accurate prediction of hypersonic flow fields over a compression ramp is critical for aerodynamic design but remains challenging due to the scarcity of experimental measurements such as velocity. This study systematically develops a data fusion framework to address this issue. In the first phase, a model trained solely on Computational Fluid Dynamics (CFD) data establishes a baseline for flow field prediction. The second phase demonstrates that enriching the training with both CFD and experimental data significantly enhances predictive accuracy: errors in pressure and density are reduced to 12.6% and 7.4%, respectively. This model also captures key flow features such as separation and reattachment shocks more distinctly. Physical analyses based on this improved model, including investigations into ramp angle effects and global stability analysis, confirm its utility for efficient design applications. In the third phase, a pre-trained model (using only CFD data) is successfully fine-tuned with experimental schlieren images, effectively reconstructing velocity fields and validating the transferability of the approach. This step-wise methodology demonstrates the effectiveness of combining simulation and experiment by pre-training and fine-tuning, offering a robust and efficient pathway for hypersonic flow modeling in real-world.
Understanding Hypersonic Flow Challenges
Designing hypersonic vehicles (flying at Mach 5+) requires an intimate understanding of how air moves over surfaces, particularly at "compression ramps"—the corners and angles where wings meet bodies. These areas are hotspots for Shock-Wave/Boundary-Layer Interaction (SWBLI), a complex phenomenon that creates separation bubbles, intense heat, and structural vibration.
Engineers usually have two tools, both flawed:
- Computational Fluid Dynamics (CFD): High-fidelity simulations are incredibly accurate but computationally "expensive," often taking hours or days for a single case.
- Wind Tunnel Experiments: These provide "ground truth" but are limited. While we can easily take Schlieren images (which visualize density gradients) or use pressure sensors, measuring the actual velocity field in a hypersonic flow is a nightmare because traditional methods struggle with extreme heat and high speeds.
Methodology: A Three-Phase Data Fusion Framework
The authors propose a three-phase data fusion framework using deep learning to combine the breadth of CFD with the physical reality of experiments.
Model Architectures and Training
Researchers compared two architectures: an Encoder-Decoder Vision Transformer (ViT) and an Attention-based CNN. To handle the specific geometry of the ramp, they used a coordinate transformation to map the physical space into a uniform computational grid. The Attention CNN was found to be superior for fluid dynamics as it better maintains the spatial continuity essential for smooth fluid gradients.
The framework utilizes Pre-training & Fine-tuning:
- Phase 1: Train a model on a large CFD dataset (112 combinations of Mach number, Reynolds number, and ramp angles) to learn general fluid physics.
- Phase 2: Enrich training with experimental measurements (pressure and density) from Mach 4 and Mach 6 Ludwieg tube wind tunnels.
- Phase 3: Fine-tune the model using experimental Schlieren images to account for real-world effects like surface roughness and freestream disturbances that CFD often ignores.
Key Results and Insights
- Error Reduction: Fusing experimental data reduced pressure prediction errors by 12.6% and density errors by 7.4% compared to CFD-only models.
- Feature Capture: The hybrid model was significantly better at identifying the exact location of the separation bubble and reattachment shocks, which are critical for calculating heat loads.
- The "Schlieren-to-Velocity" Shortcut: One of the most impressive feats was using Schlieren images—which only show density changes—to reconstruct the full velocity field. The model learned the underlying physics well enough to "fill in the blanks" that experimental hardware cannot currently see.
- Stability Analysis: The model’s predictions were accurate enough to be used for Global Stability Analysis (GSA), matching the results of expensive DNS simulations at a fraction of the computational cost.
Conclusion
While minor local discrepancies remain in very complex interaction regions, this step-wise methodology demonstrates the effectiveness of combining simulation and experiment. Future work will focus on "physics-informed" loss functions to improve Transformer continuity and expanding the framework to complex 3D vehicle shapes.
Summary updated