Neural fields for rapid aircraft aerodynamics simulations.

Area of Science:

• Computational Fluid Dynamics
• Machine Learning
• Scientific Computing

Background:

• Traditional fluid dynamics simulations are computationally expensive.
• Developing accurate and generalizable surrogate models is crucial for accelerating design and analysis.

Purpose of the Study:

• To introduce Implicit Neural Representations (INRs) for learning surrogate models of steady-state fluid dynamics.
• To demonstrate the models' ability to handle unstructured domains, geometric variations, and generalize to unseen shapes.
• To evaluate the trade-off between computational cost and accuracy.

Main Methods:

• Utilizing a coordinate-based formulation for INR-based surrogate modeling.
• Applying the methodology to datasets of 2D compressible flow over airfoils and 3D wing surface pressure distributions.
• Comparing performance against state-of-the-art Graph Neural Network architectures.

Main Results:

• Achieved over three times lower test error and improved generalization on unseen geometries compared to Graph Neural Networks.
• Demonstrated robustness to discretization and an excellent balance between computational cost and accuracy.
• Achieved a five-order-of-magnitude speedup for high-fidelity numerical solvers on a RANS transonic airfoil dataset.

Conclusions:

• INR-based surrogate models offer a powerful and efficient approach for fluid dynamics simulations.
• The proposed method significantly enhances generalization capabilities and computational speed.
• This methodology holds promise for accelerating engineering design and analysis in fluid dynamics.