Physics-Aware ML Models for Pipeline Leak Detection and Quantification

Leak detection in multiphase pipeline systems remains a critical yet challenging task due to complex flow dynamics, transient operationaldisturbances, and the scarcity of comprehensive real-world data. Traditional physics-based models, though interpretable, often exhibit reducedaccuracy under real-world uncertainties, while purely data-driven methods struggle with physical consistency and limited generalizability. Inthis study, we introduce a hybrid Physics-Informed Neural Network (PINN) framework designed for integrated leak detection, localization, andcontinuous quantification in multiphase (oil-gas-water) pipelines. By embedding mass and momentum conservation laws directly into the neuralnetwork as soft constraints, the proposed approach seamlessly combines the physical interpretability and robustness of mechanistic models withthe adaptability and predictive power of deep learning techniques. To evaluate performance, we generated high-fidelity synthetic datasetscovering diverse leak scenarios, including leak magnitudes ranging from 0.1% to 5% of nominal flow, various spatial locations, and multiplemultiphase flow regimes. The PINN framework achieved a detection accuracy of 95%, with precision and recall rates of approximately 91%,an average localization error near 5 meters, and a leak rate estimation error close to 1%. Comparative analyses against black-box neural networksand Real-Time Transient Modeling (RTTM) baselines highlight the superior performance and enhanced generalization capability of theproposed method across both classification and regression tasks. These results underscore the potential of PINNs as reliable, physicallyconsistent, and operationally actionable solutions for next-generation leak detection in complex pipeline networks, establishing a foundation forfuture large-scale field implementation and regulatory adoption.