A Study of Physics-Informed Neural Networks applied to Real-Time Atmospheric Flow Estimation for Glider Flight Planning
Keywords
Physics-Informed Neural Networks (PINNs), Atmospheric Flow Estimation, Glider Flight Planning, Real-Time Flow Prediction, Flow Field Reconstruction, Data-Driven Modeling
Presenter Abstract
Real time estimation of local atmospheric flow can be useful for gliders to improve flight performance. Traditional methods for atmospheric state estimation rely on costly sensor arrays, computationally expensive CFD simulations, or estimations based on experience. All of these approaches can be successful, but present challenges. An emerging methodology from the Machine learning (ML) front involves Physics-Informed Neural Networks (PINNs), which are a data efficient machine learning methods that reconstruct the flow field (velocity, pressure, and turbulence) from a combination of constraints from the governing equations combined with sensor data. Through combining physical governing equations with sparse measurements, the neural network (NN), or other ML method, predicts the flow field. This approach is proposed and evaluated in the context of a real time, physics consistent flow field predictions for in-flight implementation.
The framework integrates glider sensor data, such as airspeed, pressure, and angle of attack, with a neural network trained using the incompressible Navier–Stokes equations as pointwise physical constraints at collocation points generated using Latin Hypercube Sampling (LHS). Preliminary two dimensional simulations demonstrate that the proposed PINN framework can successfully reconstruct continuous velocity and pressure fields across various obstacle geometries using sparse training data, capturing overall flow behavior and maintaining physical consistency. These initial results show that PINNs can reconstruct physically reasonable flow fields, offering a promising alternative to CFD or traditional sensor based methods, with ongoing work focused on further validation and extension to three dimensional cases.
Presentations
Presented in Session 1: Platform Development I
Streaming Media
A Study of Physics-Informed Neural Networks applied to Real-Time Atmospheric Flow Estimation for Glider Flight Planning
Real time estimation of local atmospheric flow can be useful for gliders to improve flight performance. Traditional methods for atmospheric state estimation rely on costly sensor arrays, computationally expensive CFD simulations, or estimations based on experience. All of these approaches can be successful, but present challenges. An emerging methodology from the Machine learning (ML) front involves Physics-Informed Neural Networks (PINNs), which are a data efficient machine learning methods that reconstruct the flow field (velocity, pressure, and turbulence) from a combination of constraints from the governing equations combined with sensor data. Through combining physical governing equations with sparse measurements, the neural network (NN), or other ML method, predicts the flow field. This approach is proposed and evaluated in the context of a real time, physics consistent flow field predictions for in-flight implementation.
The framework integrates glider sensor data, such as airspeed, pressure, and angle of attack, with a neural network trained using the incompressible Navier–Stokes equations as pointwise physical constraints at collocation points generated using Latin Hypercube Sampling (LHS). Preliminary two dimensional simulations demonstrate that the proposed PINN framework can successfully reconstruct continuous velocity and pressure fields across various obstacle geometries using sparse training data, capturing overall flow behavior and maintaining physical consistency. These initial results show that PINNs can reconstruct physically reasonable flow fields, offering a promising alternative to CFD or traditional sensor based methods, with ongoing work focused on further validation and extension to three dimensional cases.
Presenter Biography (Optional)
Sweety Sarker is a Ph.D. student in Aerospace Engineering at Embry-Riddle Aeronautical University. Her research focuses on Physics-Informed Neural Networks (PINNs) for modeling fluid flows. She works on data-efficient flow field reconstruction from sparse measurements, with applications in atmospheric flow estimation and glider flight performance.