Pipe flow reconstruction using a novel physics-informed machine-learning technique

A novel method for improving pipe leak detections is numerically investigated through a staged

acoustic analysis. Unsteady acoustic wave propagations within the fluid, during a hypothetical

leak scenario, are modelled using coupled Large-Eddy Simulations (LES) and Acoustic

Perturbation Equations (APE) in OpenFOAM. The base case represents a three-dimensional pipe

with circular leaks centered along the axial direction of the pipe. The Reynolds number

and leak diameter are varied in the range of Re=25000–100000 and D=0.8-4.0 mm, respectively.

Preliminary results suggest that the generation of acoustic sources coincide with a

velocity jet and low-pressure zone within the leak region.

A reduced-order model framework has been developed to reconstruct flow field variables

using sparse acoustic pressure measurements obtained from probes along the pipe surface. The

approach incorporates a classifier to localize potential leak regions, while physics-informed

neural networks (PINNs) are employed to enforce physical consistency through the residual

minimization of the continuity and momentum equations. The hybrid data-driven and physics-

constrained strategy demonstrates strong potential for reducing false-positive leak detections

caused by non-leak anomalies, such as internal blockages, surface erosion, or structural

vibrations. Ongoing extensions of the framework include generalization to more complex pipe

configurations, such as those featuring orifice plates and curved sections.