Noise Removal Tool for Compressible and Incompressible Flows Using a Physics-Informed Neural Network

Noise removal from data or sensors is a crucial challenge in experimental studies, where traditional techniques such as filters and smoothers are commonly employed. However, these methods often lack a physics-based foundation, requiring either domain expertise or an extensive trial-and-error process to achieve satisfactory outcomes. Moreover, their performance deteriorates as noise levels increase, often leading to significant distortion of the original signal or derived quantity based on the signal. Physics-informed neural networks (PINNs) provide an innovative solution by integrating physical laws and governing equations into the machine learning framework. This study investigates the potential of PINNs to enhance noise removal compared to conventional numerical methods. Specifically, we assess the performance of numerical methods both with and without filtering alongside PINNs in the context of transport equations. We have studied the effectiveness of the above-mentioned techniques on synthetically generated subsonic and supersonic flows from numerical simulation. The results reveal that PINNs can effectively reconstruct pressure information from noisy velocity data, a task that traditional numerical methods do not perform adequately on noisy high Reynolds number cases. This finding highlights the superior capability of PINNs in addressing noise-related challenges in signal processing, particularly under high-noise conditions and high Reynolds number cases.