The risk of CO2 and brine leakage to environmental receptors is one of the main concerns associated with geologic CO2 storage. Legacy wells associated with oil and gas activities are abundant and may be found within the area of review of some storage projects. If legacy wells are present within a project area, they will require constant surveillance to ensure they are properly completed and plugged to prevent fluid transmission. Hence, deployment of an efficient monitoring system for early detection of CO2 and brine leakage from failed wells is imperative to mitigate environmental and financial risks. In this project, we developed an efficient and cost-effective near-surface monitoring package, tested in the field and combined with machine learning, that can provide real-time and long-term surveillance of plugged and abandoned (P&A) wells.

We developed an advanced and cost-effective monitoring technique to detect and locate potential CO2 leakage signals in real time through anomalies observed in natural environmental variability without any need to historical baseline surveys. The proposed monitoring technique aims to introduce the field implementation of such a baseline-free monitoring technology applicable to near-surface leakage detection originating form P&A wells. The proposed monitoring system can be employed to any projects which involve the use of the subsurface with the risk of unintended leakage such as carbon capture and storage, fluid disposal, unconventional resource development, and geothermal energy projects.

Summary of Research Outcomes
We conducted pilot and field-scale water and CO2-controlled release experiments through a proxy well, with sentinel surveillance conducted by a modular monitoring package consisting of soil sensors and a weather station. The soil sensor package measured the response of soil conditions [e.g., soil water content (WC), temperature, and electrical conductivity (EC)] to the leakage. The experiments were designed to artificially simulate point-source leakage of fluids through a proxy well from a depth of around 3 ft, the typical cut-off depth for surface casing. The experimental apparatus consists of a proxy well casing and annulus installed in the center of a trench (which was then backfilled) with dimensions of 6 ft×6 ft and depth of 5 ft. We released fluids (CO2 and water) at a depth of 5 ft (2 ft below the deepest sensor array) in the wellbore annulus behind the casing. Soil signatures (e.g., soil WC, temperature, and EC) as well as temperatures inside the casing were continuously monitored during and after fluid release. Various experiments involving the release of fluids at different leakage rates and durations were tested. It is found that soil EC was found to be the most sensitive soil signature to CO2 leakage. The leaked CO2 increased the soil's electrical conductivity (EC), with the extent of the increase depending on the initial soil water content. We further developed a machine learning model for early CO2 leakage detection by discerning anomaly patterns related to leakage in the monitoring data. The proposed ML framework leverages commonly measured meteorological data, alongside soil WC and EC. These inputs were used to predict a binary label indicating anomalous observations, whether caused by CO2 leakage or natural environmental variations. To enrich the training dataset, we first employed a Physics-Informed Neural Network (PINN) to estimate soil WC and EC through incorporating constraints from Richard’s equation, along with meteorological data and CO2 leakage rates. We tested the PINN by dividing the dataset into separate training and testing datasets, ensuring that the test set included diverse meteorological and CO2 injection scenarios. After training and validating the model, we generated synthetic datasets simulating soil water content and electrical conductivity under various CO2 injection rates and meteorological conditions. Once the dataset was enriched, we implemented a feedforward Bayesian Neural Network to predict the probability of an observation being anomalous. The ML results showed that combining field data and the PINN model can enhance the performance of the neural network by accurately separating leakage response from natural variations in soil.

Summary of Project Impacts
We developed an advanced and cost-effective monitoring technique to detect and locate potential CO2 leakage signals in real time through anomalies observed in natural environmental variability without any need to historical baseline surveys. The proposed monitoring technique aims to introduce the field implementation of such a baseline-free monitoring technology applicable to near-surface leakage detection originating form P&A wells. The proposed monitoring system can be employed to any projects which involve the use of the subsurface with the risk of unintended leakage such as carbon capture and storage, fluid disposal, unconventional resource development, and geothermal energy projects.