Physics-Informed Machine Learning Advisory System for Corrosion Prediction in Gas Producer Wells

Abstract Gas producer wells experience complex internal corrosion mechanisms accelerated by multi-phase flow regimes, produced-water chemistry, and thermal or pressure cycling. Missed or late detection cascades into well-integrity impairment, unplanned shutdowns, and, in extreme cases, catastrophic loss of containment. While corrosion modelling has been studied and developed in the literature for flow in pipeline, corrosion modelling is less understood and modelled for flow inside the wells. There arises a need to better understand governing corrosion factors and predict corrosion rates with higher confidence. This paper presents a physics-based machine-learning corrosion advisory system designed for gas wells. The objective is to predict maximum metal loss (MML) per well using routinely available data: historical Multi-finger Imaging Tool (MIT) logs, production and wellhead operating data, formation and produced-fluid chemistry, and multiphase-flow dynamics. The advisory system predicts corrosion rates, alerts for high-risk wells, and advises inspection and intervention scheduling. Accordingly, the operator will be able to have higher visibility of well-integrity risks and act preemptively reducing surveillance costs, mitigating deferred production and alleviating high-consequence failures. The corrosion-prediction problem is treated as a supervised machine-learning task. Data sources include MIT inspection logs, well chemical profiles, multiphase-flow characteristics, and operational data trends from field gauges. Random Forest (RF) and k-Nearest Neighbors (KNN) models were trained to predict MML per well. Feature-importance and Principal Component Analysis (PCA) techniques were used to identify the most informative variables and enhance model performance. The models were validated on held-out wells and demonstrated high predictive accuracy, achieving a mean absolute error below 0.015, and ±5 % absolute wall-thickness deviation on blind tests. Field application to a sample of gas wells confirmed the model's ability to identify severe-corrosion cases and revealed new operational and chemical drivers of internal corrosion. The system delivers actionable insights that improve maintenance prioritization, mitigate hazardous conditions, minimize unplanned shutdowns, and reduce inspection cost. Integration into digital dashboards enables near-real-time corrosion forecasting. This work introduces a practical, physics-informed machine-learning tool that accurately predicts current and 12-month-ahead MML using existing field data and predicted production. It represents a novel step toward real-time, data-driven corrosion management in gas operations, offering measurable benefits in safety, cost savings, and predictive maintenance planning for practicing engineers. The solution approach presented in this paper has been filed as a patent in US Patent Office on Aug 23rd, 2025, under #IS24.1764-US-NP