Development of a Physics-Informed Neural Network for Aerodynamic Analysis of Airfoils and Wings

July 7, 2026

Computational Fluid Dynamics (CFD) simulations remain, to this day, very computationally intensive and restrict a CFD engineer’s freedom, as they require simplifications to be made in most problems so that they can be solved within a reasonable time frame. Additionally, CFD results demand significant memory requirements and are challenging to post-process. A Neural Network (NN) can help alleviate the computational intensity and memory constraints. Instead of storing all CFD data, a neural network learns an approximation of the solution (velocity and pressure fields) to the numeric models of the Navier-Stokes (NS) equations used by the CFD solvers after being trained using a dataset of CFD simulations. Once trained, the final model is a relatively small function that is capable of computing the flow variables at any given point of the flow in an instant, regardless of whether the input data matches a point in the training dataset or not. A Physics-Informed Neural Network (PINN) extends this approach by incorporating the residuals of the governing equations in order to learn the underlying physics, which improves the physical fidelity of the solution over a default NN. Additionally, a PINN has the freedom to cheaply apply methods like adaptive mesh refinement that are very computationally intensive in traditional CFD. The proposed PINN framework is capable of producing results with low errors on top of the errors of the CFD simulations, while only taking a fraction of the time and space complexity.

Authors in the community:

• 

  R

  Rúben Soares Lopes

  ist199554

Supervisors of this institution:

• 

  F

  Frederico Afonso

  ist156309

• 

  F

  Filipe Szolnoky Ramos Pinto Cunha

  ist14018

mechanical-engineering - Mechanical engineering

eng - English

Instituto Superior Técnico