A Data Fusion Framework for Grounding Aerospace Surrogate Model via Experimental Wind-Tunnel Observations
Researchers have developed a data fusion framework that enhances the predictive accuracy of aerospace surrogate models by integrating experimental wind-tunnel observations with computational fluid dynamics (CFD) simulations.
Intelligence analysis by Gemini 2.5 Flash
This paper introduces a novel correction framework designed to adapt deep learning surrogate models, initially trained on high-fidelity CFD data, using real-world wind-tunnel pressure-sensitive paint (PSP) measurements. The goal is to overcome systematic discrepancies between simulated and experimental aerodynamic predictions, thereby improving the reliability of AI models in aerospac…
Imagine you have a super-smart computer program that designs airplane wings, but it only learns from other computer programs. This paper teaches that program to also look at real-world tests in a wind tunnel, like a giant fan blowing air over a model plane. By doing this, the computer program gets much better at predicting how a real wing will fly, making future airplanes safer and more efficient.
Analysis
The development of accurate aerodynamic surrogate models is a critical area in aerospace engineering, offering the potential to significantly reduce the computational cost and time associated with traditional design processes. While deep learning surrogates trained on high-fidelity Computational Fluid Dynamics (CFD) data can accurately reproduce numerical predictions for both scalar outputs and entire fields, their real-world predictive fidelity is often hampered by systematic discrepancies when compared to actual experimental observations. This paper addresses this fundamental challenge by proposing an experimentally grounded correction framework.
Geotransolver Surrogate
The core of this research involves a Geotransolver surrogate model, which was initially trained on an extensive dataset of 2,300 high-fidelity CFD simulations. These simulations covered the NASA CRM wing-body configuration, incorporating variations in geometry, Mach numbers ranging from 0.70 to 0.85, and angles of attack from 0 to 4 degrees. While this CFD-trained surrogate demonstrated exceptional accuracy in reproducing CFD-integrated aerodynamic forces and pitching moments, achieving an R2 value greater than 0.99, it notably failed to align with experimental data, highlighting the persistent gap between simulation and reality.
PSP Measurements
To bridge this gap, the researchers introduced a correction network that leverages wind-tunnel Pressure-Sensitive Paint (PSP) measurements. This innovative approach allows for the incorporation of experimental information without the need to retrain the entire, computationally intensive surrogate model. The correction network was specifically trained on spatially registered PSP measurements collected at two freestream Mach numbers (0.70 and 0.85) across the same angle-of-attack range. Its primary function is to learn the systematic discrepancy between the surrogate-predicted and experimentally measured surface-pressure distributions, effectively acting as a fine-tuning layer.
NASA CRM Wing-Body
The application of this framework to the NASA CRM wing-body configuration yielded significant improvements. At Mach 0.85, the correction substantially enhanced agreement with the PSP measurements, particularly in critical areas such as the wing suction peak, shock location, and subsequent pressure recovery. This resulted in a reduction in both the magnitude of the prediction error and the fraction of the wetted surface where the error exceeded 0.05 in Cp. Crucially, this improvement was achieved using a limited experimental dataset and without altering the parameters of the pretrained surrogate. On held-out angles of attack, the grounded surrogate demonstrated agreement with measurements to within 2.3-2.7% of the measured Cp range, outperforming direct interpolation between measured conditions at every state tested. This validates the framework's ability to ground large-scale simulation-trained surrogates by learning CFD-to-experiment discrepancies while preserving generalization capability and computational efficiency.
Key points
- Aerodynamic surrogate models trained on CFD data often show systematic discrepancies with experimental observations.
- A new data fusion framework uses wind-tunnel Pressure-Sensitive Paint (PSP) measurements to correct CFD-trained deep learning surrogates.
- The Geotransolver surrogate, trained on 2,300 CFD simulations of the NASA CRM wing-body, was significantly improved.
- The correction network learns discrepancies without retraining the original surrogate, preserving its computational efficiency.
- The grounded surrogate demonstrated substantially improved agreement with experimental measurements, outperforming direct interpolation.
This framework could significantly reduce the need for extensive physical wind-tunnel testing by providing highly accurate, experimentally-grounded surrogate models, leading to faster and more cost-effective aerospace design cycles. It also enhances the reliability of AI-driven design tools.
The effectiveness of this correction framework relies heavily on the quality and availability of experimental data. Limited or noisy wind-tunnel measurements could hinder the correction network's ability to accurately learn discrepancies, potentially leading to models that are still not fully reliable for real-world applications.
