Project Title: Predictive Surrogate Modeling of Aerodynamic Characteristics for a 2D NACA 0012 Airfoil using CFD and Machine Learning
Date of Presentation: 25/10/2026
Team - YantraAi
Team member :-
Daniyal Shaikh
Gaurav S. Vidhate
Prachi S. Vishwasrav
Sanjana Kashyap
One sentence Description :-
Delivering instant aerodynamic predictions without the need for a supercomputer.
The Specific Problem
Computational fluid dynamics requires massive supercomputing power. Solving complex physical equations turns simple design iterations into massive bottlenecks that can take days to process.
Who Faces This?
Aerospace teams designing efficient aircraft wings and automotive engineers trying to extend the battery range of electric vehicles are constantly slowed down by these testing limits.
Current Limitations
Today's workflows rely entirely on slow, expensive server farms running trial-and-error tests. Engineers simply do not have a way to validate their ideas in real-time
Why It Matters
By shifting from heavy physics solvers to an AI surrogate, we reduced evaluation time from 36 hours down to just 5 seconds, saving massive amounts of money and carbon emissions.
The Engineering Bottleneck
1. User Input
The engineer simply types their target parameters into the Agent37 dashboard.
Parameters:
2. The AI Engine
Instead of building a mesh and running a heavy physics simulation, the system immediately processes the inputs.
It routes the data through our pre-trained machine learning weights to instantly map the aerodynamic forces.
3. Instant Results
In under 5 seconds, the dashboard renders the exact performance metrics on the screen.
Outputs:
• Coefficient of Lift (Cl)
• Coefficient of Drag (Cd)
Solution
We created an AI-driven surrogate model that acts as a lightning-fast digital twin for aerodynamic testing.
To train it, we ran 100 high-fidelity CFD simulations using a smart space-filling statistical method.
We strictly bounded our training data to angles of attack under 12.09 degrees.
This brilliant constraint kept the physics clean and completely eliminated the chaotic noise of deep aerodynamic stall.
Process
Solution and process
CFD vs. Experimental (NASA TM-4074) shows excellent agreement in the steady-state regime (α ≤ 12.09°).
R² = 0.999 for Lift Validation, proving exceptional accuracy.
Beyond α = 12.09°, CFD predicts massive flow separation (resulting in lift drop and drag spikes).
ML model successfully validated against experimental and CFD data across design points (α = 0° to 16°).
ML model accurately traces non-linear lift and drag trends at higher angles of attack.
Key Takeaway: CFD + ML integration provides high-fidelity performance estimation with excellent predictive behavior.
Result and Validation
Instead of feeding messy data into our AI, we strictly bounded our training set to the stable, attached-flow regime to prevent the model from learning chaotic deep-stall noise.
1. Smart Data Bounding
Traditional fluid solvers iterate complex physics equations over millions of grid cells. Our surrogate model bypasses the grid entirely, mapping the exact same physics instantly.
2. Instant Over Iterative
Our Deep Neural Network completely failed on our 100-run dataset, producing negative R² values. We solved this by pivoting to classical models that are far more robust for small datasets.
3. The Overfitting Trap
Ingenuity & Technical Decisions
4. Rejecting Deep Learning
We chose a simple Linear Regression model over complex Neural Networks because it perfectly mirrored the linear physics of attached airflow without structural overfitting.
Potential & Next Steps
Beyond The Hackathon
Aerospace startups, electric vehicle designers, and renewable energy teams can use this tool to optimize their designs without needing a massive computing budget.
The Real-World Value
It democratizes aerodynamic engineering by replacing expensive, multi-million-dollar supercomputer hours with an instant, free web dashboard.
Future Upgrades
Our immediate next step is to expand the AI's training dataset to handle full 3D wing geometries and predict turbulent wakes in extreme deep-stall conditions.
Path To Production
To become a true industry standard, we need to build API plugins that feed these AI predictions directly into CAD software like SolidWorks as the engineer draws.