Abstract
The U.S. Environmental Protection Agency requires replacement of all lead service lines within 10 years, yet accurately identifying buried lead pipes remains a major challenge. Existing detection methods, such as excavations, electrochemical sensors, or ground-penetrating radar, are often expensive, disruptive, or sensitive to environmental noise. We present a noninvasive approach that combines physics-based finite element analysis (FEA) surrogate modeling with machine learning (ML) to detect lead pipes efficiently. A computationally efficient FEA model was developed to simulate the dynamic behavior of buried copper and lead pipes, incorporating key features such as stop-valve openings to enable realistic loading conditions. Transient dynamic simulations analyzed mechanical responses, specifically pipe acceleration, under varying geometries and loading scenarios. Over 13,000 synthetic observations were generated, with added noise and signal masking to reflect real-world sensor limitations. Seven ML models were trained on these acceleration signals to classify pipe material. K-nearest neighbor (KNN) and Extreme Gradient Boosting (XGBoost) achieved the highest performance, each reaching 99.9% classification accuracy. This integrated modeling and ML framework offers a scalable, cost-effective method for utilities to locate and replace lead pipes, supporting regulatory compliance while minimizing operational disruptions and resource expenditures.