Physics-informed multi-task neural network for simultaneous leak detection, diameter estimation, and localization in multiphase pipelines

Pipeline leaks are loss-of-containment incidents with potentially severe process-safety consequences fire, explosion, toxic release, and environmental damage yet their simultaneous detection, diameter estimation, and localization in offshore multiphase systems remain unsolved due to nonlinear gas-liquid flow dynamics and sparse subsea instrumentation. Existing approaches address these tasks separately, relying either on computationally intensive real-time transient models or on purely data-driven detectors that lack physical consistency, leaving a gap in the physics-consistent characterization required as input to dynamic risk assessment and operational risk management frameworks. This paper presents a Physics-Corrected Enhanced Physics-Informed Neural Network (PCE-PINN), a unified multi-task architecture that simultaneously performs leak detection, diameter estimation, and location prediction from standard inlet/outlet pressure and flow-rate measurements alone. The framework couples conservation-law-derived features with an embedded physics subnetwork that penalizes violations of nondimensional mass, momentum, and mechanical-energy balance through adaptive curriculum training and R adaptive collocation. Evaluated via stratified 5-fold cross-validation on a mechanistically generated dataset representative of a laboratory test facility at Hamad Bin Khalifa University (HBKU), PCE-PINN attains F1 = 1.00 for leak detection together with strong regression performance: R2 = 0.906 for leak diameter and R2 = 0.941 for normalized leak location. Against five classical machine learning baselines on identical features, PCE-PINN improves location R2 by 16.5% over the best baseline, confirming that gains originate from the physics-informed architecture. SHAP-based attribution confirms that predictions are governed by physically interpretable, conservation-consistent features. By supplying simultaneous, physics-consistent estimates of leak presence, severity, and location from routine process measurements, PCE-PINN provides the diagnostic layer required to support dynamic and operational pipeline risk management in offshore multiphase systems.