AI-Driven Computational Fluid Dynamic Simulations and Experiments to Predict Methane Emission: Applications to Idle and Abandoned Wells

Abstract Emission from idle and abandoned wells can pose significant safety risks because of methane leakage. Many of such wells in the United States are in residential and commercial areas, with some locations unknown due to the drilling before the establishment of oil and gas regulations. Each year, idle wells add to the number of abandoned wells, potentially developing integrity issues over time. These issues may arise from compromised barriers such as cement cracks, cement-formation bond cracks, and cement-casing bond cracks. Given the sheer number of idle wells and the flammable nature of methane, there is a need for real-time monitoring solutions with minimal electrical-mechanical components. Our proposed canopy system integrates sensors and communication kits with Internet of Things (IoT) capabilities, enabling operators to remotely monitor methane leaks in real-time. Artificial intelligence enhances the capabilities of our canopy system. Because the physical mechanisms of methane leaks in abandoned wells remain inadequately understood, we approach the problem from a surface-level perspective. A digital twin can be created based on the unique statistical profile of a well's emissions. Canopies deployed temporarily collect live data, which is used to train neural networks to predict future leaks, thereby reducing risks and minimizing the physical footprint of the canopies. Machine learning algorithms can also translate concentration time series data into mass leak rates. While numerical simulations are time-intensive and impractical for extreme conditions, AI models trained on simulation data are more efficient. This paper explores novel machine learning techniques for simulations and live data predictions.