Efficient Aircraft Design Optimization Using Multi-Fidelity Models and Multi-fidelity Physics Informed Neural Networks

Imagine you are an architect trying to design the perfect airplane. To make sure the wings won't snap and the plane will fly efficiently, you need to run thousands of computer simulations.

The Problem: The "Slow Motion" Simulation
Traditionally, these simulations are like trying to paint a masterpiece using only a tiny, fine-tipped brush. You get incredibly accurate results (the "High-Fidelity" data), but it takes forever. If you want to test 100 different wing shapes, you might be stuck waiting months for the computer to finish the calculations.

On the other hand, you could use a big, chunky brush (the "Low-Fidelity" model). It's super fast and cheap, but the picture looks blurry and misses the tiny details. You can test 100 shapes in a day, but you don't know if the design is actually safe or efficient.

The Solution: The "Smart Translator"
This paper introduces a clever new method called Multi-fidelity Physics-Informed Neural Networks (MPINN). Think of this not as a new brush, but as a super-smart translator or a magic lens.

Here is how it works, using a simple analogy:

1. The Two Maps

Imagine you have two maps of the same city:

The Low-Fidelity Map: A rough sketch on a napkin. It shows the main roads and the general shape of the city, but it's missing side streets and details. It's easy to draw and quick to look at.
The High-Fidelity Map: A satellite image with every single tree, house, and pothole visible. It's perfect, but it's huge and hard to process.

2. The "Magic Lens" (The AI)

The researchers built an AI (a type of computer brain) that acts as a Magic Lens.

You feed the AI the rough napkin sketch (the fast, low-quality data).
The AI has already studied a few satellite images (the expensive, high-quality data) and knows the laws of physics (how wind and pressure actually work).
Using what it learned, the AI looks at your rough sketch and instantly "fills in the blanks." It predicts what the satellite image would look like for that specific sketch, adding all the missing details without you having to run the expensive simulation.

3. How the AI Learns (The Recipe)

The paper explains that the AI doesn't just guess; it uses a special recipe to combine two types of corrections:

The Linear Correction: This is like saying, "Okay, the rough sketch is 10% too small, so let's just stretch it a bit." It handles the simple, predictable differences.
The Non-Linear Correction: This is the fancy part. It says, "Wait, the wind doesn't just stretch; it swirls and creates complex patterns here." This part of the AI learns the tricky, messy physics that the rough sketch missed.

By combining these two, the AI can take a cheap, fast simulation and turn it into a result that looks almost exactly like the expensive, slow one.

Why This Matters

In the past, if you wanted to test a new airplane wing, you had to wait days for the computer to crunch the numbers.

Before: 1 design test = 1 week of waiting.
With this new method: You can test 1,000 designs in the time it used to take to test 1.

The AI learns the "rules of the game" (physics) and uses them to upgrade the cheap, blurry pictures into high-definition masterpieces instantly.

The Future: The "Auto-Designer"

The paper suggests that in the future, we won't just use this to predict results. We could combine it with other AI tools (like GANs, which are like AI artists that can draw new things) to create a self-driving design process.

Imagine a system where you say, "I want a wing that is lighter and flies faster," and the AI automatically draws hundreds of new wing shapes, instantly predicts how they would perform using its "Magic Lens," and picks the winner—all in a matter of minutes, not months.

In a nutshell: This research is about teaching computers to be smart enough to use cheap, fast guesses to create accurate, high-quality results, saving the aerospace industry massive amounts of time and money.

1. Problem Statement

Aircraft design optimization traditionally relies on high-fidelity simulation techniques such as the Finite Element Method (FEM) and Finite Volume Method (FVM). While these methods provide accurate results, they are computationally expensive and time-consuming, particularly when fine meshes are required. This creates a bottleneck in Multidisciplinary Design Optimization (MDO), where rapid iteration through multiple design alternatives is necessary.

Existing solutions, such as surrogate models (e.g., polynomial regression, Radial Basis Functions) and Reduced Order Models (ROMs), often struggle with two main limitations:

Scalar vs. Field Quantities: Most existing surrogate models predict scalar outputs (e.g., total lift, drag, weight) rather than full field quantities (e.g., stress distribution, displacement fields across a grid), which are critical for detailed design analysis.
Grid Discrepancies: Handling differences in grid topology between low-fidelity (coarse) and high-fidelity (fine) simulations is challenging. Traditional manifold alignment methods (MA-ROM) often fail when the discrepancy between datasets is large or when the data exhibits complex, non-linear features requiring many modes to represent accurately.

2. Methodology

The research proposes a Multi-fidelity Physics-Informed Neural Network (MPINN) framework to predict high-fidelity field variables from low-fidelity data, leveraging the physical laws governing the system to bridge the fidelity gap.

Theoretical Framework

The core relationship between low-fidelity ( $y_{LF}$ ) and high-fidelity ( $y_{HF}$ ) data is modeled as a function $F(x, y_{LF})$ . The paper adopts a decomposition strategy proposed by Meng and Karniadakis, separating the correlation into linear and non-linear components:
$\hat{y}_{HF} = \alpha F_l(x, y_{LF}) + (1 - \alpha) F_{nl}(x, y_{LF})$
Where:

$\alpha$ is a hyper-parameter controlling the degree of nonlinearity.
$F_l$ and $F_{nl}$ represent linear and non-linear correlation functions, respectively.

Neural Network Architecture

The MPINN architecture consists of three fully connected neural networks:

NNL (Low-Fidelity Network): Approximates the low-fidelity data using spatial coordinates as input. It serves as the baseline prediction.
NNH1 (Linear Correction): Estimates the linear correlation between low and high-fidelity data.
NNH2 (Non-linear Correction): Estimates the non-linear correlation, utilizing L2 regularization to capture complex relationships.

The final high-fidelity prediction is the sum of the low-fidelity prediction and the corrections provided by NNH1 and NNH2. The model is trained by minimizing a loss function that includes the Mean Squared Error (MSE) between predicted and true values for both fidelity levels.

Implementation Case Study

Task: Predicting absolute pressure fields for a 2D NACA 2412 airfoil under 10 m/s airflow.
Data Generation: Simulations were run using ANSYS Fluent.
- Low-Fidelity: 870 nodes.
- High-Fidelity: 21,630 nodes.
Preprocessing: Data was aligned by resampling high-fidelity pressure values onto the low-fidelity grid using the griddata interpolation method to ensure consistency during training.

3. Key Contributions

Field Quantity Prediction: Unlike traditional surrogate models that focus on scalar outputs, this approach successfully predicts full field quantities (pressure distribution) across different grid resolutions.
Non-Linear Fidelity Bridging: The implementation of MPINN allows for the modeling of complex, non-linear relationships between coarse and fine meshes, overcoming the limitations of linear manifold alignment methods.
Physics-Informed Efficiency: By integrating physics constraints (via the neural network structure and loss functions), the model achieves high accuracy with significantly fewer high-fidelity training samples, which are computationally expensive to generate.
Proof-of-Concept Validation: The study demonstrates the feasibility of using MPINN to approximate high-fidelity results from low-fidelity inputs with minimal computational overhead.

4. Results

The proof-of-concept task successfully demonstrated the model's capability:

Accuracy: The MPINN model effectively bridged the gap between the 870-node low-fidelity dataset and the 21,630-node high-fidelity dataset.
Visualization: Comparative analysis (Fig. 3 in the paper) showed that the predicted high-fidelity pressure fields closely matched the actual high-fidelity simulation results, validating the model's ability to reconstruct fine-grid details from coarse-grid inputs.
Efficiency: The approach significantly reduces the time and cost associated with running full high-fidelity simulations for every design iteration.

5. Significance and Future Outlook

This research offers a transformative pathway for Multidisciplinary Design Optimization (MDO) in the aerospace industry:

Accelerated Design Cycles: By enabling rapid, accurate evaluation of design alternatives without running expensive FEM/FVM simulations for every iteration, the design process can be significantly sped up.
Automation Potential: The success of MPINN lays the groundwork for integrating advanced tools like Generative Adversarial Networks (GANs) and Manifold-aware ROMs to create fully automated design workflows. These future systems could continuously refine aircraft geometries based on performance feedback.
Cost Reduction: Reducing reliance on high-fidelity simulations lowers the computational cost of aircraft development, making the optimization process more accessible and efficient.

Conclusion: The paper establishes that multi-fidelity machine learning, specifically MPINN, is a robust solution for overcoming the computational bottlenecks in aircraft design. It successfully balances accuracy and efficiency, paving the way for a future where aircraft design is driven by automated, physics-informed, and data-efficient optimization loops.