Multiphase System Leak Detection, Localization, and Dimension Estimation using Multi-Head Physics Informed Neural Network

Accurate and real-time leak detection in multiphase subsea pipelines remains a challenging problem due to highly nonlinear flow interactions, sparse sensor availability, and the limited generalization capability of traditional data-driven and CFD-based approaches. Existing intelligent leak detection methods often fail to incorporate the underlying physical laws, while current physics-informed models typically address only single-task predictions and lack uncertainty awareness, limiting their reliability in practical offshore environments. To address these gaps, this study proposes a Multi-Head Physics-Informed Neural Network (MPINN) that simultaneously estimates leak location, leak size, and pressure distribution while enforcing multiphase flow physics through embedded governing equations. The framework integrates a physics-consistent loss formulation, task-decoupled prediction heads, and Monte Carlo–based uncertainty quantification to enhance robustness under noisy and sparse real-world conditions. Experimental evaluation shows that MPINN achieves 100% leak classification accuracy, an R² of 95% for leak size prediction, and an R² of 96.25% for leak location estimation, significantly improving spatial characterization performance. The model also produces calibrated confidence intervals that effectively capture aleatoric and epistemic uncertainty, strengthening its reliability for risk-informed decision-making. Overall, this work introduces a unified, physics-aware, and uncertainty-informed framework that advances the state of the art in intelligent multiphase pipeline monitoring and provides a practical foundation for future industrial deployment.