Leveraging Physics-Informed Neural Networks for Noise Reduction in Fluid Flow Data

Noise removal from sensor data remains a significant challenge in experimental studies. Traditional approaches such as filters and smoothers are widely used but often lack a physics-based foundation. These methods typically require domain expertise and extensive trial-and-error tuning, and their effectiveness diminishes with increasing noise levels—frequently resulting in distortion of the original signal or derived quantities. Physics-Informed Neural Networks (PINNs) offer a promising alternative by embedding physical laws and governing equations directly into the learning process. In this study, we evaluate the potential of PINNs for noise removal in transport equations and compare their performance with conventional numerical methods, both with and without filtering. Our analysis focuses on synthetically generated subsonic and supersonic flow data from numerical simulations. Results demonstrate that PINNs can successfully reconstruct pressure fields from noisy velocity data, outperforming traditional methods, especially in high Reynolds number scenarios. Conventional approaches struggle to denoise velocity fields in supersonic flows and often yield nonphysical pressure distributions. To address this, we introduce an adaptive-weight and adaptive-viscosity PINN framework that enables robust pressure reconstruction in supersonic regimes. These findings underscore the superior capability of physics-informed models in handling noise, particularly in compressible flows with shocks, where traditional filters fail. This study highlights that physics-based filtering may be essential for accurate reconstruction in shock-dominated problems.