Machine Learning-Driven Quantification of CO₂ Plume Dynamics at Illinois Basin Decatur Project Sites Using Microseismic Data

Abstract

This study utilizes machine learning to quantify CO₂ plume extents by analyzing microseismic data from the Illinois Basin Decatur Project (IBDP). Leveraging a unique dataset of well logs, microseismic records, and CO₂ injection metrics, this work aims to predict the temporal evolution of subsurface CO₂ saturation plumes. The findings illustrate that machine learning can predict plume dynamics, revealing vertical clustering of microseismic events over distinct time periods within certain proximities to the injection well, consistent with an invasion percolation model. The buoyant CO₂ plume partially trapped within sandstone intervals periodically breaches localized barriers or baffles, which act as leaky seals and impede vertical migration until buoyancy overcomes gravity and capillary forces, leading to breakthroughs along vertical zones of weakness. Between different unsupervised clustering techniques, K-Means and DBSCAN were applied and analyzed in detail, where K-means outperformed DBSCAN in this specific study by indicating the combination of the highest Silhouette Score and the lowest Davies–Bouldin Index. The predictive capability of machine learning models in quantifying CO₂ saturation plume extension is significant for real-time monitoring and management of CO₂ sequestration sites. The models exhibit high accuracy, validated against physical models and injection data from the IBDP, reinforcing the viability of CO₂ geological sequestration as a climate change mitigation strategy and enhancing advanced tools for safe management of these operations.

1. Introduction

The escalating concerns surrounding climate change and the associated rise in atmospheric carbon dioxide (CO₂) levels necessitate the exploration of effective carbon capture and storage (CCS) solutions. The increasing concentration of CO₂ and other greenhouse gasses in the atmosphere is a primary driver of global warming, leading to severe environmental consequences such as melting polar ice caps, rising sea levels, and more frequent and intense weather events [1]. As nations strive to meet international climate targets set by agreements such as the Paris Agreement, there is an urgent need to develop and implement technologies that can significantly reduce CO₂ emissions from industrial sources [2].

Geological sequestration of CO₂, involving the injection of CO₂ into deep underground formations, has emerged as a promising strategy to mitigate greenhouse gas emissions [3,4]. This process entails capturing CO₂ from emission sources, transporting it to a suitable geological site, and injecting it into deep rock formations such as depleted oil and gas fields, deep saline aquifers, or unmineable coal seams [3]. The CO₂ is stored securely in these formations, preventing its release into the atmosphere and thus contributing to the reduction in overall atmospheric CO₂ levels [5]. This method not only aids in mitigating climate change but also enhances the viability of fossil fuel-based industries by providing a pathway to reduce their carbon footprint.

The Illinois Basin Decatur Project (IBDP) stands at the forefront of such initiatives, offering a comprehensive dataset that includes well logs, microseismic activity records, and CO₂ injection metrics [6]. Located in the Illinois Basin, the IBDP is one of the leading CCS demonstration projects globally, aiming to advance the understanding and implementation of CO₂ sequestration technologies [7]. The project’s extensive dataset is invaluable for research and development in this field, providing detailed insights into the behavior of CO₂ in the subsurface environment [8]. This dataset includes continuous monitoring of CO₂ injection processes, capturing critical data on how the CO₂ plume evolves over time within the geological formation [9].

In this study, we explore the application of machine learning methodologies to quantify the spatial extent of CO₂ plumes by leveraging microseismic data and injection data obtained from the IBDP site, spanning from November 2011 to June 2018. Machine learning, a subset of artificial intelligence, involves the development of algorithms that can learn from and make predictions based on data [10]. By applying these advanced computational techniques to microseismic data from the IBDP, this research aims to accurately map and predict the movement of CO₂ within the storage formation. Microseismic monitoring, which records the small-scale seismic events induced by CO₂ injection, provides crucial information about the subsurface movements and the behavior of the injected CO₂ [11]. The integration of machine learning with microseismic data analysis offers a powerful tool for enhancing the precision and reliability of CO₂ plume monitoring and prediction [12,13].

Previous interpretations of microseismic events at the Illinois Basin Decatur Project (IBDP) site have varied. Some show events predominately in Mount Simon [14], while other interpretations show events concentrated in the basement [15]. It appears that it is unclear if the datum is the measured depth from the surface or if it is the sea level, although the data are reported as measured depth not subsea levels, suggesting that the data were recorded from the local surface (~675 feet). However, we believe the best interpretation aligns the observed data with fundamental physical principles like buoyancy effects. Aligning the majority of detected seismic events within the basement rather than the reservoir contradicts the expected behavior of buoyant CO₂ displacing denser brine [16]. In similar projects, such as the Sleipner site, the expected buoyancy CO₂ migration creates distinct seismic signatures in the subsurface layers [13,17].

To address these discrepancies, this study employs machine learning to reinvestigate and more accurately allocate CO₂ saturation using seismic, geological, and operational data, including baffles, localized barriers, and injection logs. The primary goal is to enhance the accuracy and consistency of subsurface interpretations by integrating various data sources with advanced computational techniques.

For this study, we conducted a comprehensive well log analysis to identify baffles and fractures, utilizing both traditional well log techniques and advanced methods involving computer vision and deep learning [12]. Additionally, we explored the correlation between microseismic events and the densities of baffles and fractures, complemented by Pulsed Neutron Logging (PNL) to assess CO₂ saturation distribution. We then analyzed the clustering and dynamics of microseismic events, both 2D and 3D, over time using unsupervised machine learning techniques, correlating these findings with three key data sources: fracture and baffle intensities, microseismic activity, and PNL-derived CO₂ saturation distribution. Our analysis, supported by an invasion percolation model for CO₂ migration and fundamental physical principles such as buoyancy effects, provides a coherent explanation of the observed CO₂ plume dynamics and the corresponding microseismic signatures.

2. Materials and Methods

2.1. Illinois Basin Decatur Project Site

The Illinois Basin Decatur Project (IBDP), a key initiative of the Midwest Geological Sequestration Consortium (MGSC) and funded by the U.S. Department of Energy, successfully proved the potential for CO₂ injection and storage in geological formations. The project’s primary goal was to inject 1 million tonnes of CO₂, sourced from Archer Daniels Midland Company (ADM) in Decatur, IL, USA, into the Mt. Simon Sandstone at a depth of approximately 2100 m.

The Mt. Simon Sandstone, located within the Illinois Basin—a 155,000 square kilometer cratonic basin—is a significant saline reservoir with an estimated storage capacity ranging from 11 to 150 billion tonnes. Early geophysical studies conducted in 2007 confirmed the location’s suitability by detecting no faults that would render it unsuitable for CO₂ storage [6].

The Mount Simon Sandstone, from the Cambrian period, is a vast and thick sandstone layer found extensively in the Midwestern United States, especially within the Illinois Basin. Its extensive regional spread and significant thickness make it an excellent choice for CO₂ storage, offering a substantial and stable reservoir. At the Illinois Basin Decatur Project (IBDP) site, the Mount Simon Sandstone exceeds 457 m (1500 feet) in thickness. The upper section was formed in a tidally influenced environment, while the lower 183 m (600 feet) is composed of arkosic sandstone from a high-energy braided river or alluvial fan system, which provides excellent secondary porosity due to the dissolution of feldspar grains [18].

The Eau Claire Formation, which lies above the Mount Simon Sandstone, serves as the primary caprock. It is 212 m (695 feet) thick and mainly consists of shale with layers of fine-grained sandstones. Due to its low permeability and porosity, it acts as an effective barrier, preventing the upward migration of CO₂ and ensuring the long-term stability and safety of the storage site. Together, the Mount Simon Sandstone and the Eau Claire Formation create a robust geological sequestration system, which is crucial for the success of the IBDP [19].

In November 2011, CO₂ injection started after test wells were drilled and a compression plant was built. A variety of methods are used to closely monitor the injected CO₂ to guarantee the efficiency and safety of the carbon storage operation. To make sure the CO₂ stays confined, pressure monitoring is employed to track the movement and distribution of the CO₂ within the storage reservoir, ensuring that it remains contained. In addition to microseismic monitoring, 3D seismic monitoring attempted to detect any potential leaks or movement of the CO₂ underground [6]. Figure 1 shows The Illinois Basin Decatur Project site along with the location of major project elements. The red outline for the demonstration site is approximately 800 m (2600 ft) on the side. CCS1 is the injection well and VW1 is the monitoring well. The orange line is the CO₂ pipeline extended to the compressors, shown as orange rectangles in the southeast corner of the image.

2.2. Subsurface Supercritical CO₂ Transport and Storage Mechanisms

Supercritical CO₂ is essential for carbon capture and storage (CCS) due to its liquid-like density and gas-like viscosity, enabling effective penetration and storage volumes in porous rock formations. Achieved at temperatures above 31.04 °C and pressures above 73.82 bar, it enhances fluid displacement and storage efficiency in deep saline aquifers and depleted reservoirs [20]. Temperature and pressure variations during CO₂ injection influence its phase and interaction with formation water, affecting migration and distribution. These factors are critical for machine learning models analyzing microseismic data to map CO₂ plumes and identify leakage pathways.

The fluid front is the interface between injected CO₂ and resident fluids within the storage reservoir. Monitoring its progression helps calibrate models and detect anomalies, indicating potential storage risks or inefficiencies. Factors such as reservoir permeability, porosity, and pressure conditions influence the fluid front’s movement. As CO₂ is injected, it displaces resident fluids, creating a moving front through porous rock formations.

Understanding fluid front behavior is crucial for effective CO₂ storage [21]. A uniform and controlled fluid front maximizes reservoir utilization and prevents bypassing low-permeability zones, which can lead to inefficient storage and leakage risks. The heterogeneous nature of subsurface geology, including variations in rock permeability and the presence of faults and fractures, creates complex flow patterns affecting CO₂ distribution.

Injection rate and pressure also play critical roles in shaping the fluid front [22]. High injection rates can advance the fluid front rapidly, risking low-permeability zone bypass and caprock fracturing, while controlled rates ensure uniform fluid displacement and enhanced storage security. Additionally, temperature differences between injected CO₂ and formation fluids create thermal gradients that influence CO₂ viscosity and density, impacting fluid front stability and behavior.

The buoyancy effect plays a critical role in CO₂ behavior and stability during carbon capture and storage (CCS). Since supercritical CO₂ is less dense than resident fluids such as formation brine, it rises and accumulates beneath caprock due to buoyant forces, despite gravity and capillary forces attempting to retain it [23]. In the Illinois Basin Decatur Project, supercritical CO₂’s lower density compared to brine causes it to rise and spread toward the top of the Mt. Simon Sandstone, creating a distinct CO₂–brine interface and facilitating CO₂ migration to higher reservoir regions. This movement is influenced by reservoir permeability, porosity, and geological structures, which can redirect or impede CO₂ flow. Microseismic monitoring tracks this buoyant migration, revealing areas of increased pressure and potential fracturing.

2.3. Importance of Natural Fractures and Baffles

Natural fractures and baffles at the Illinois Basin Decatur Project (IBDP) play a crucial role in understanding subsurface dynamics and optimizing CO₂ sequestration. Baffles can impede vertical movement and concentrate lateral spreading of the CO₂ front within a higher permeability thief zones. While fractures can enhance storage by increasing the contact area for processes like dissolution and mineral trapping, they also pose risks by providing potential pathways for CO₂ leakage if they connect to overlying or adjacent permeable layers.

We have implemented advanced machine learning techniques, including computer vision and deep learning, and significantly improved the identification and analysis of these subsurface features. Within the measured depth interval, from 5371.5 feet to 7198 feet, of the Mount Simon Sandstone, 5875 baffles were identified [12]. As expected in a vertical well through an essentially flat lying cratonic formation, numerous fractures were not recognized. The numerous baffles highlight the potential influence of baffles and localized seals on the upward and lateral movement of buoyant CO₂. Our results suggest an invasion percolation model for upward CO₂ migration, where the CO₂ plume trapped within sandstone intervals as highly saturated layers periodically breaks through barriers by overcoming gravity and capillary forces.

2.4. Microseismic Monitoring

Microseismic events, which are small-scale seismic activities, play a crucial role in monitoring subsurface dynamics in carbon capture and storage (CCS) projects [24]. These events provide essential data on the geomechanical responses of storage reservoirs to CO₂ injection, helping to ensure site integrity and stability. During CO₂ injection, increased pressure can induce microseismicity by altering stress distributions and potentially reactivating faults or fractures [22]. Sensitive seismic sensors detect these events, enabling precise localization and characterization [25]. Microseismic event location for monitoring CO₂ injection can perturb the ambient stress field and trigger brittle failure on small fractures or faults [11,26]. These data help operators understand CO₂ plume dynamics and maintain the structural integrity of storage sites.

2.5. Machine Learning in Seismic Data Analysis

Machine learning, especially unsupervised learning techniques such as clustering, is increasingly vital in geological studies and seismic data analysis [27,28,29,30]. These advanced computational methods uncover patterns and features within seismic data that traditional methods might miss, enhancing the ability to discern underlying geological structures and anomalies. Among these techniques, K-Means clustering is widely used [31], partitioning data into predefined clusters based on inherent features, which facilitates more accurate and insightful interpretations of complex datasets. Another valuable algorithm is DBSCAN (Density-Based Spatial Clustering of Applications with Noise), which identifies clusters of varying shapes and sizes while distinguishing noise from significant data points [32]. In this study, both K-Means and DBSCAN clustering techniques are optimized and implemented, and the outcomes are compared using the Silhouette Score and Davies–Bouldin Index.

We have also developed an opensource tool to visualize and analyze microseismic data using both K-Means and DBSCAN clustering techniques. This tool, accessible through an online dashboard, utilizes the Python programming language and packages such as NumPy and Pandas, offering data loading and processing functionalities, and is available at https://cz63dzvzquchsikhdpugsq.streamlit.app/ accessed on 8 July 2024. The dashboard accounts for varying time spans to detect CO₂ plume pressure dynamics both laterally and vertically from the injection point. Additionally, the dashboard offers the option to optimize the K-Means and DBSCAN clustering techniques. Once the parameters are set, the workflow is applied to both datasets provided: “SLB Data" and "Relocated Data,”.

2.5.1. Data Collection and Visualization

We have used the dataset spanning from December 2011 to July 2018, which provides detailed information on microseismic events related to CO₂ injection at the Illinois Basin Decatur Project (IBDP). The Mt. Simon Sandstone layers (A Lower, A Upper, B, C, and D) serve as the primary reservoir for CO₂ injection, offering significant storage capacity due to their porosity and permeability. The dataset, provided in the Supplementary Materials, contains two sheets, "SLB Data" and "Relocated Data," each detailing microseismic events, including the following:

• Origin Time: exact date and time of each microseismic event.
• Horizontal Difference: horizontal displacement in feet from a reference point.
• Depth Difference: vertical displacement in feet.
• Total Difference: combined horizontal and vertical displacements.
• Year/Mo. Category: categorizes events by year and month for temporal analysis.

2.5.2. Data Preprocessing and Feature Engineering

The initial step in this study involved comprehensive preprocessing of microseismic data to ensure reliability and accuracy for subsequent analyses. This included data cleaning, normalization, and feature extraction. Data cleaning addressed missing, incomplete, or erroneous entries, enhancing data quality. Normalization adjusted the data to a standard scale, ensuring equal contribution of all features in the clustering process. The key features considered were SLB Depth Difference (vertical displacement), SLB Horizontal Difference (horizontal movement), time of occurrence, and geological formations. Standardizing these features prevented any single feature from dominating the clustering process, allowing for balanced consideration of all seismic event aspects. This resulted in a well-prepared dataset for machine learning analysis, effectively contributing to identifying patterns and anomalies in CO₂ plume dynamics.

2.5.3. Data Visualization

Data visualization modules were developed using Python and Streamlit to enable the visualization of depth difference data for selected datasets. The process involved selecting the time period of the investigation, selecting the data base, and choosing chart type.

2.5.4. Unsupervised Clustering Techniques

Unsupervised clustering techniques are essential tools in data analysis, allowing for the discovery of inherent patterns and structures within unlabeled datasets. Among the most widely used methods are K-Means and DBSCAN (Density-Based Spatial Clustering of Applications with Noise). The K-Means clustering section involves applying the K-Means algorithm to group similar data points in the dataset based on features like depth and horizontal differences. This unsupervised learning technique helps identify clusters of seismic line segments with similar characteristics, revealing potential geological formations or patterns within the seismic data. K-Means works by partitioning the data into a predefined number of clusters without prior knowledge of the labels

The DBSCAN clustering utilizes Density-Based Spatial Clustering of Applications with Noise (DBSCAN) on the dataset. This unsupervised learning technique identifies clusters based on density, effectively pinpointing areas with high microseismic activity and detecting outliers. Unlike K-Means, DBSCAN does not require pre-specifying the number of clusters. Instead, it uses two key parameters: epsilon (eps) and min_samples. Epsilon defines the radius around each point to assess density, while min_samples sets the minimum number of points within epsilon to form a dense region. DBSCAN is more effective in identifying clusters of arbitrary shapes and varying sizes, a flexibility not offered by K-Means, which assumes spherical clusters of similar sizes. By integrating both techniques, K-Means provides an initial partitioning of clusters, while DBSCAN refines these clusters by uncovering intricate patterns and isolating outliers. The Silhouette Score and Density-Based Clustering Validation (DBCV) Index are also used to evaluate the quality of the clustering and density separation between clusters [33].

3. Results

Clustering the microseismic events by time intervals revealed vertical distributions within restricted ranges away from the injection well. Within these clusters, the upward distance from the injection zone displayed a complex distribution of event frequencies and amplitudes. This pattern supports a model of periodic CO₂ migration, where buoyant CO₂ accumulates in thin, highly saturated layers within sandstone intervals. These layers periodically overcome gravity and capillary forces, breaching discrete barriers or baffles and allowing CO₂ to migrate vertically through zones of weakness.

The following sections detail the application of machine learning clustering techniques and their validation with physical models, along with conventional analysis of microseismic events and well logs.

3.1. Microseismic Monitoring of IBDP

At the Illinois Basin Decatur Project (IBDP), a comprehensive monitoring system, including geophones in both the injection well CCS#1 and the nearby monitoring well VW#1, provides extensive coverage [6]. The analysis of microseismic data from the IBDP reveals the temporal and spatial distribution of microseismic events, offering insights into the CO₂ plume’s subsurface behavior. This monitoring system recorded 10,123 microseismic events, indicating CO₂ movement and pressure buildup, particularly in the Mt. Simon A formation in CCS#1 and the entire Mt. Simon column in VW#1. Figure 2 shows the map of the IBDP site with all microseismic events relative to the CCS1 injection well and VW1 monitoring well from December 2011 to February 2018 (gray), with events from January 2012 to 30 April 2012 highlighted. Advanced techniques like FMI logging identified baffles that significantly influenced CO₂ migration, often halting its upward movement near the injection point CCS#1 and causing lateral movement away from CCS#1, leading to localized pressure increases. From December 2011 to February 2012, data showed a clear pattern of horizontal spread of these events away from the injection point. This observation, in line with previous studies [12], indicates the movement of the CO₂ plume laterally away from the injection point, rather than an upward movement.

From April 2012 to the end of the injection in December 2014, microseismic events were concentrated near the injection perforations, indicating pressure buildup around the injection point. This led to the CO₂ plume breaching the baffles and expanding vertically upward. At this stage, the lateral expansion of the CO₂ plume was confined to less than 4000 ft, as opposed to the earlier extension around 9000 ft away from the injection point.

Figure 3 shows these events spreading both vertically (ranging from −500 ft to 1100 ft) and horizontally (extending up to 9000 ft), suggesting the CO₂ plume’s extension and interaction with various geological layers, including the Mt. Simon Sandstone and the Precambrian basement. Initially, the CO₂ plume spread laterally away from the injection point, accumulating around 9000 ft away (December 2011–February 2012). The baffle density log on the right margin of Figure 3 shows the presence of high-density baffles above the injection zone that facilitate these horizontal movements. It then moved back toward the injection point during March 2012 and concentrated around the injection point for the rest of the injection time and post-injection period until February 2018. As pressure builds up near the injection point, it overcomes gravity and capillary forces, resulting in breaching discrete barriers or baffles and allowing CO₂ to migrate vertically through zones of weakness. Post-injection events mainly occurred between 2600 and 3900 ft away from the injection point, extending up to lower Mount Simon C.

3.2. Pulsed Neutron Logging (PNL) for CO₂ Saturation Distribution

Pulsed Neutron Logging (PNL) is a crucial technique for monitoring CO₂ saturation and formation fluids near the wellbore. PNL logs provide comparative analyses by contrasting pre-injection (base pass) measurements with subsequent monitoring logs taken during and after the injection period in both the CCS#1 injection well and the VW#1 monitoring well. A time-lapse evaluation of PNL data indicated that most of the CO₂ was contained within the lower Mt. Simon A interval, with minor extension into the Upper Mt. Simon A in the CCS#1 injection well. Post-injection PNL logs revealed minimal changes in CO₂ saturation, except for minor buoyancy effects that reduced saturation at the bottom of the permeable zones in CCS#1. However, repeat PNL logs of VW1 highlighted vertical CO₂ plume movement up to Mt. Simon D and Mt. Simon. This vertical movement is attributed to the presence of highly dense baffles identified using FMI logs in Mt. Simon A, which prevent vertical CO₂ plume movement in the CCS#1 well and redirect the CO₂ plume toward VW#1 (Figure 4). The pressure buildup in VW#1 leads to CO₂ periodically breaching these localized barriers or baffles and migrating upward to Mt. Simon D and Mt. Simon. The PNL log was obtained during four runs: in 2014, 2016, 2018, and 2019. These intervals allowed us to monitor the dynamic changes in CO₂ distribution over time. The data reveal that CO₂ reached the VW#1 monitoring well and migrated through the Mt. Simon formation, extending up to the Eau Claire Shale caprock. The main concentration of CO₂ was observed in both the Lower and Upper Mt. Simon intervals. Interestingly, CO₂ was not detected throughout the Mt. Simon formation in the injection well, CCS#1, where, during all five years of monitoring, the CO₂ remained contained within the Lower and Upper Mt. Simon intervals. These observations provide critical insights into how barriers and baffles direct the CO₂ plume towards the VW monitoring well, where it encounters vertical fractures and subsequently moves upward. Our findings are consistent with earlier studies by [34], which demonstrated that CO₂ plume geometry is strongly influenced by low-porosity and low-permeability layers (baffles), significantly affecting CO₂ distribution at distances greater than 984 feet from CCS#1. Similar phenomena were also observed in the CO₂ sequestration project at Citronelle Dome, Alabama [35].

3.3. K-Means Clustering Techniques

K-Means partitions data into a predefined number of clusters without prior knowledge of the labels. To determine the optimal number of clusters (k), methods such as Total Variation and Within-Cluster Sum of Squares (WCSS) are used. Total Variation measures the overall spread within clusters, while WCSS calculates the sum of squared distances between data points and their cluster centers. The optimal k is identified at the point where adding more clusters results in diminishing returns, often visualized using an "elbow curve." This curve bends at the optimal point, indicating that further increases in k provide only marginal improvements.

In this case, three clusters were identified as optimal by WCSS, while Total Variation suggested five clusters, as shown in Figure 5. To refine the selection, both three and five clusters were evaluated using the Silhouette Score and Davies–Bouldin Index. The Silhouette Score measures how similar a point is to its own cluster compared to other clusters, ranging from −1 to 1, with higher values indicating better-defined clusters. The Davies–Bouldin Index evaluates the average similarity ratio of each cluster with its most similar cluster, where lower values indicate better clustering.

Table 1. Comparison of Silhouette Score and Davies–Bouldin Index of 3 and 5 clusters identified as optimum number of clusters using WCSS and Total Variation techniques.

Number of Clusters	Silhouette Score	Davies–Bouldin Index
3	0.683	0.427
5	0.623	0.473

summarizes the Silhouette Score and Davies–Bouldin Index for K-Means clustering, indicating that three clusters have a higher Silhouette Score and a lower Davies–Bouldin Index for the SLB dataset.

Figure 6 illustrates the K-Means clustering of microseismic events in a 3-dimensional space of time, horizontal difference, and depth difference with respect to the injection point using three clusters. Initially, the CO₂ plume moved laterally away from the injection point and vertically up to Mt. Simon C. Over time, it began to accumulate around the injection point and did not extend beyond 4000 ft laterally from the injection point.

To better understand CO₂ plume dynamics, dimensionality reduction to 2D subsets was performed, and K-Means clustering was applied for visualization. This required new clustering, resulting in different clusters based on the chosen dimensions. Figure 7 shows the determination of the optimal number of clusters when the 3D original data are mapped onto a 2D plane of depth difference and horizontal distance from the injection point. Similar to the original dataset, the "elbow curve" for WCSS suggested three clusters, while Total Variation suggested five clusters.

Table 2. Comparison of Silhouette Score and Davies–Bouldin Index of 3 and 5 clusters identified as optimum number of clusters using WCSS and Total Variation techniques for 2D subset of data mapped on depth difference–horizontal difference distance to the injection point.

Number of Clusters	Silhouette Score	Davies–Bouldin Index
3	0.623	0.464
5	0.651	0.355

compares these indices for three and five clusters, confirming that five clusters are optimal for this subset of data.

The optimal number of clusters for each subset was determined using “elbow curves” for both WCSS and Total Variation and confirmed by the highest Silhouette Score and the lowest Davies–Bouldin Index.

Table 3. Silhouette Score and Davies–Bouldin Index analysis of different 2D subsets of the original dataset.

Data 2D Subsets	Silhouette Score	Davies–Bouldin Index
SLB Depth Difference vs SLB Horizontal Difference	0.651	0.355
Time vs SLB Horizontal Difference	0.638	0.461
SLB Depth Difference vs Time	0.638	0.461

compares these indices across different subsets, confirming that the physical interpretation of clusters aligns with PNL logging observations and FMI log data on baffle intensity, supporting five clusters as the optimal number (Figure 4).

Figure 8 illustrates the K-Means clustering of microseismic events on the 2D plane of depth difference and horizontal distance from the injection point, highlighting the horizontal and vertical extension of the CO₂ plume. The y-axis represents the depth difference relative to the injection point, while the x-axis shows the horizontal spread of the events.

Five clusters of microseismic events were identified. Clusters 0, 1, and 2, which had lower event densities, occurred early in the injection period (December 2011 to early March 2012) and were located farther from the injection point. This suggests an initial horizontal spread of CO₂ due to baffle confinement, followed by upward movement as pressure built up, leading to breakthroughs near the injection point. From mid-June 2012 to the end of injection in November 2014, and continuing post-injection until February 2018, CO₂ accumulated around the injection point.

Figure 9 displays the clustering of microseismic events in a 2D subset of time (in 10^6 s) versus horizontal distance from the injection point. The optimal number of clusters for this subset was five. Initially, microseismic events spread up to 9000 feet from the injection point but later concentrated closer. Post-injection activities primarily occurred between 2600 and 3900 feet away, with fewer events near the injection point. This pattern suggests that high-density baffles and barriers near the injection point closed after injection stopped and pressure decreased below the closure pressure.

Figure 10 shows microseismic events in the 2D subset of depth difference versus time. The optimal number of clusters here was also five. This clustering reveals vertical CO₂ movement throughout the injection and post-injection periods. Being lighter than formation water, CO₂ accumulated in highly saturated sandstone layers, overcame gravity and capillary forces, and breached barriers to migrate vertically. Most microseismic events were concentrated in Mt. Simon A upper and Mt. Simon B, with fewer events in Mt. Simon A lower and some extending to Mt. Simon C and the Precambrian.

3.4. DBSCAN Clustering Techniques

DBSCAN uses two key parameters: epsilon (eps) and min_samples. Epsilon defines the radius around each point within which density is assessed, while min_samples sets the minimum number of points required within this radius to form a dense region. The choice of the "eps" value significantly influences DBSCAN clustering.

A common technique to determine the optimal "eps" value is the K-Distance graph. This graph plots the distances of each point to its k-th nearest neighbor. A sharp change in the graph indicates a good candidate for the "eps" value. Figure 11 shows that the optimal "eps" value is 0.2, according to the K-Distance graph for the 5th nearest neighbor distance.

DBSCAN clustering of the SLB whole dataset suggests six major clusters and one outlier cluster, presented as cluster −1, as shown in Figure 12.

Table 4. Silhouette Score of DBSCAN clustering.

Epsilon	Number of Clusters	Silhouette Score
0.2	7	0.338

compares the Silhouette Score and Davies–Bouldin Index for DBSCAN and K-Means clustering. The results indicate that K-Means clustering outperforms DBSCAN and will therefore be used for the final conclusions.

4. Discussion

The Illinois Basin Decatur Project (IBDP) provided an extensive dataset for analyzing CO₂ plume dynamics, leveraging microseismic event monitoring through a comprehensive system including geophones in both the injection well CCS#1 and the nearby monitoring well VW#1. The collected data, spanning from November 2011 to December 2014, comprised 10,123 recorded microseismic events. The analysis of these events, particularly using advanced machine learning (ML) clustering techniques like K-Means and DBSCAN, has yielded significant insights into the subsurface behavior of the CO₂ plume, especially in the context of identifying and understanding the role of geological features such as baffles and local barriers.

Initially, spatial and temporal K-Means clustering revealed distinct vertical and horizontal distributions of microseismic events, indicative of the CO₂ plume’s movement. Clusters identified early in the injection period (December 2011 to early March 2012) displayed a lateral spread up to 9000 ft away from the injection point, as CO₂ encountered baffles that confined its vertical migration. Subsequent pressure buildup led to breakthroughs, allowing upward migration. This phenomenon was evident from mid-June 2012 to the end of the injection and post-injection period (February 2018), where microseismic events concentrated near the injection point, highlighting the cyclical nature of CO₂ migration and accumulation. The clustering of microseismic events in multi-dimensional and 2D spaces, respectively, confirmed the dynamics observed. The optimal clustering configuration identified five clusters, as validated by both Silhouette Scores and Davies–Bouldin Indices, underscoring the robustness of the K-Means approach. This clustering aligned well with physical observations from Formation Micro-Imager (FMI) logs, which identified high-density baffles and barriers, corroborating the ML-derived patterns of CO₂ movement. The FMI logs showed cemented or shale-rich intra-formational intervals, acting as leaky seals inhibiting vertical migration, a critical factor influencing the plume’s horizontal and vertical dynamics. The PNL log data further supported these interpretations, detailing the pressure dynamics around the injection point.

DBSCAN clustering, while effective in identifying major and outlier clusters, as shown in Figure 10 and Figure 11, did not perform as well as K-Means in providing clear, actionable insights. The effectiveness of K-Means clustering in interpreting CO₂ plume dynamics is evident in its ability to delineate clusters that align with physical geological features and monitoring logs. This alignment not only validates the clustering results but also enhances the understanding of CO₂ behavior in the subsurface, informing better management and prediction strategies for carbon sequestration projects.

5. Conclusions

The application of machine learning clustering techniques, particularly K-Means, has significantly enhanced the interpretation of CO₂ plume dynamics at the Illinois Basin Decatur Project (IBDP). By effectively delineating spatial and temporal patterns of microseismic events, K-Means clustering provided crucial insights into the horizontal and vertical movements of the CO₂ plume, especially in relation to geological features such as baffles and localized barriers. The clustering results, validated by FMI and PNL log data, highlighted the initial lateral spread followed by vertical migration due to pressure build-up and the post-injection stabilization. Comparative analysis demonstrated K-Means’ superior performance over DBSCAN, affirming its robustness for geospatial analysis. This integration of ML techniques with physical monitoring enhances the accuracy of subsurface modeling, informing better carbon sequestration strategies and ensuring long-term project efficacy. Integrating machine learning with FMI image logs and PNL log monitoring has significant implications for CO₂ capture, utilization, and storage (CCUS). It enhances the precision of CO₂ plume pressure and saturation analysis during (area of review) and after injection (post injection site care), which improves site management. This integration helps in assessing legal pore space ownership, ensuring cap rock integrity, and evaluating reservoir capacity, thereby optimizing regulatory compliance, operational safety, and overall project efficiency. These advancements contribute to more effective carbon sequestration strategies and support progress toward climate change mitigation goals.