Characterization of a fluvial deltaic salt-influenced system for CO₂ storage using seismic poststack inversion and self-organizing maps in the upper Miocene, Onshore Louisiana Open Access

The United States is the second largest carbon emitter after China, and within the country, the southeast region emits the most CO₂ (Meckel et al., 2021). However, southern states such as Louisiana contain large subsurface trends of high reservoir quality sands that could sequester carbon dioxide and nearby infrastructure necessary for commercial scale projects (Zulqarnain et al., 2023). These reservoirs have high net to gross sands and wide-spread seals necessary for containment. In addition, the local geology comprises shale-sand heterogeneity, which can greatly improve carbon storage integrity (Langhi et al., 2021; Pourmalek et al., 2021). With these ideal formations, it becomes critical to know the potential carbon storage that is available. This research focuses on high porosity Miocene sand trends in south Louisiana and shows a new workflow for identifying carbon injection targets and estimating the pore space volume.

The workflow leverages two different methodologies: self-organized mapping (SOM) and poststack acoustic impedance inversion. SOM is an unsupervised machine-learning technique that takes multiple seismic inputs and clusters the information into a 2D map (Kohonen, 1982). To better see the 2D map projection, the color bars can be manipulated, thus providing geologic details that interpreters may otherwise miss (Matos et al., 2010). To improve mapping consistency, edits were made to the temporal space to avoid mixing geology in 2D maps. Amplitude data and AVO attributes have been used as inputs for SOM to identify direct hydrocarbon indicators in class 2 and 3 settings (Roden and Chen, 2017). Recent applications of SOM show its usefulness for delineating seismic facies changes for conventional, unconventional, and even gas hydrate targets (Strecker and Uden, 2002; Chopra et al., 2021; La Marca and Bedle, 2022; Lubo-Robles et al., 2023). Early results also suggest that SOM can highlight relative differences in hydrocarbon saturations for prospects found within the same survey (Chenin and Bedle, 2022).

Poststack inversion has been implemented for years by geophysicists to estimate the earth’s acoustic impedance. The method combines seismic reflectivity with estimated wavelet, noise, and geologic response. There are three general approaches used to compute poststack inversion: sparse-spike, model-based, and recursive. Previous studies show that a deterministic model-based inversion provides a better result to differentiate lithology and predict porosity for real data (Almutairi et al., 2022) and further confirmed using forward modeling with synthetic data (Russell and Hampson, 1991). Low acoustic impedance anomalies can tie to high porosity lithologies that would be favorable for possible CO₂ injection (Alshuhail et al., 2009).

This research builds on the application of SOM using poststack seismic attributes and complements its results with an acoustic impedance inversion, which provides porosity information and improved resolution. By using the two methodologies together, carbon storage targets can be better delineated and characterized. The study begins by first showing the geologic background of south Louisiana and the Middle Miocene formation. Next, the paper introduces SOM technique along with acoustic impedance inversion workflow. The methods are applied to a deltaic system that is influenced by salt tectonics, and the work shows a potential project-sized carbon storage target. The paper concludes with carbon storage estimations for the target and the advantages of leveraging the techniques in this workflow.

Southeastern Louisiana is part of the Gulf Coast geologic province of the south-central United States (Figure 1) and records a geologic history from Jurassic time to present. The sediment source is from two depositional trends: the ancestral Mississippi and ancestral Tennessee fluvial-deltaic axes (Combellas-Bigott and Galloway, 2006). These trends have regular interplay of sand and shale depositions based on relative sea level rise and fall. Seismic survey coverage for the study area does not capture the full source-to-sink depositional system. Conceptually, the shelf edge is north of the survey area, and there is a shelf-edge delta system prograding to the south and into the study area. Data coverage within our zone of interest primarily images the slope systems connected to the prograding deltas. These prograding systems can produce large slope-fan systems with high porosity and high permeability sands, which is a focus of this study as a potential carbon storage target.

Sequence boundaries are key correlation markers that divide the upper and middle Miocene formation and have been implemented for mapping the Miocene trend production (Vidrine, 1971). Biostratigraphic boundaries are dated by correlation to key benthonic and planktonic foraminifera and calcareous nannoplankton datums to the Berggren et al. (1995) time scale. This research uses these key boundaries and examines the upper Miocene M3 surface through the upper section of the Miocene M7. The M5 sands contain well-imaged fluvial-deltaic lobes in the northern portion of the focus area. The deposition is dominated by prograding deltaic sands and shales near the shelf break.

Faulting is prevalent throughout the area, and key stratigraphic intervals are impacted by these structures. Structurally, complexity is due to the number of salt domes in the area (Figure 2). During the mid-Jurassic, the Louann Salt was widely deposited across the northern portion of the Gulf of Mexico (Galloway et al., 2000). The Jurassic and Cretaceous sediment loading caused the salt to undergo a wide range of salt deformations, which are further influenced by local depositional trends. Entering the Paleocene, the western Gulf of Mexico is loaded with sediment, while the northern and eastern portions are sediment starved. This differential sediment loading initiates the Terrebonne allochthonous salt sheets. As the Eocene transitions to the Miocene, Mississippi and Tennessee river systems begin to load the allochthonous salt sheets, which caused the salt to evacuate and provide further sediment accommodation. Once the Terrebone evacuation was exhausted, the deeper, autochthonous Louann Salt began to expand basinward to accommodate the high rate of sediment loading. The Miocene is generally characterized by the spreading of allochthonous salt driven by high rates of sedimentation in the Gulf of Mexico (Galloway et al., 2000). This salt tectonic has a large impact on spatial and temporal sediment deposition in south Louisiana and the local study area.

The data for this study are a combination of proprietary seismic reflection data and petroleum exploration wells. Included are a 310 mi² (803 km²) 3D seismic survey (merged and reprocessed in 2007) and seventeen wells that contain compressional sonic and gamma-ray logs. All the wells except one, the D4, are used as control wells in the seismic inversion (Figure 2). The D4 is the only well containing a measured bulk density log and is used as a blind well test. Pentolite was the original acquisition source used during the 1990s as part of the geophysical and geologic exploration of oil and gas prospects in south Louisiana. The survey has a bin size of 82.5 ft × 165 ft and an average fold of 40.

The workflow uses two techniques in conjunction: SOM and acoustic impedance inversion (Figure 3). SOM can be done relatively quickly using any combination of seismic attributes that aid in the interpretation. In comparison to SOM, acoustic impedance inversion takes more time due to the calibration of well data and building a low-frequency model that predicts an appropriate background geologic trend. This methodology views SOM as an effective screening method that can be used concurrently with inversion products.

Combining multiple seismic attributes to identify a feature of interest often can become confusing and disjointed. To combine multiple attributes into a clear image, researchers have leveraged machine learning and labeled their output as seismic facies (Roy et al., 2013). SOM is an excellent seismic facies classification tool that captures the information residing in input seismic attributes by reorganizing data samples based on their topological relation (Zhao et al., 2016). The technique’s algorithm attempts to mimic human pattern recognition and produce projected 2D maps. These maps can aid during interpretation where a single seismic attribute does not properly identify with a geologic feature.

To find potential CO₂ injection targets, seismic attributes delineate zones with rock properties suitable for CO₂ storage — namely high injectivity sands with high porosity and permeability. Previous SOM work shows that attributes such as amplitude, similarity, sweetness, spectral frequency, and curvedness can highlight depositional features that are associated with high quality sands suitable for CO₂ storage (Zhao et al., 2016). Using the attribute-assisted seismic processing and interpretation software, the analysis tested multiple attributes as SOM inputs such as dip magnitude, coherency, energy, entropy, and coherent energy. All tests were conducted with raw data attributes with no normalization or weighting of inputs. Similarity, peak frequency, and coherent energy were the final selected inputs during SOM testing. Interpreted surfaces extracted these attributes along chronostratigraphic events within the 3D seismic data set to avoid crossing geologic boundaries and blurring the signal. Extractions in this fashion better highlight fluvial depositional elements.

The analysis used standard values accepted for siliciclastic environments where $α=0.31$ and $β=0.25$⁠. Wavelet estimation is necessary to tie wells, and the well ties derived a statistical, zero-phased wavelet from the zone of interest. This wavelet provided the best well-tied correlations during testing when compared with other band-pass and trapezoidal wavelets. Wavelet extraction shows a dominant frequency of 20 Hz, which was implemented into the inversion. However, inspection of frequencies at the M5 showed higher peak frequencies averaging 30 Hz. At this interval, tuning is estimated to be 78 ft.

Interpreted horizons guide the low-frequency model (LFM) as it interpolates acoustic impedance values between well locations. Thus, it is critical to have sufficient well data available to properly constrain the lateral variability in the background trend. The sixteen control wells properly captured the low-frequency variability across the 310 square mile study area. Low frequencies missing from the seismic but calculated from well data and incorporated through the LFM are extremely important to obtain accurate impedance values that correlate to rock properties. Inversion analysis and quality control performed at various well locations shows a strong match between the inverted acoustic impedance (in red) and the computed impedance from the well (in blue) through the M3–M7 target window (Figure 5).

After reviewing and testing various models, a model-based inversion was executed for the full study area. The target window for the inversion begins at the M3 top and continues down through the upper M7. This section covers approximately 3000’ true vertical depth (914 m).

After testing several combinations of attributes as SOM input, three seismic attributes showed improved results for delineating fluvial depositional elements: similarity, peak frequency, and coherent energy (Figure 6). This combination of attributes was selected due to its ability to clearly define the deltaic lobe and channel systems. The resulting SOM seismic facies output shows geologic features derived from geophysical data and highlights a 26 mi² (67 km²) deltaic lobe feature in the northern part of the survey (Figure 6d). SOM results were corendered with variance to better highlight depositional features to assess how they fit the conceptual understanding of the depositional model. Smaller channels and lobe edges are better resolved, and the seismic facies associated with the lobe sands are mostly absent in the eastern side of the survey. SOM suggests that the western side of the area is dominated by the sand seismic facies within the zone of interest. The maps provide a moderate increase in detailed seismic geomorphologic features that are missed when only using a single attribute to map. Depositional elements identified are channels, levees, lobes, and background deposits.

Equation 2 transforms acoustic impedance data to porosity values. The porosity map and SOM results can be used together for interpretation (Figure 8). SOM suggests that sand is deposited across a large portion of the area but in contrast the porosity map shows the best quality sands are concentrated in the northern area of the survey. Gamma-ray logs across the area map the Sand 1 zone and align with the sand seismic facies that SOM produces (Figure 9). The cross section in Figure 9 shows SOM and relative acoustic impedance backdrops. The SOM output qualitatively indicates where sand increases. It has less resolution in comparison to the inversion output. Horizon interpretation and well-log correlation show structure of the lobe to be on the flank of a syncline with varying thickness along dip. Well data show an average gross reservoir thickness of 90 ft (27 m) and average net sands to be 81 ft (25 m), which is at the seismic tuning level for this area.

Porosity, derived from acoustic impedance, and SOM reveal geologic features with the potential for high amounts of geologic carbon storage. Together, the two maps can be used to better understand the lateral extent and porosity distribution of possible CO₂ containers. A comparison of the two techniques readily shows similarities. Both methods show the two northern deltaic lobes of interest and how faulting impacts the lateral distribution of the sands. The SOM maps suggest that better-quality sands are mostly found in the western side of the area. Gamma-ray comparison to SOM suggests that the final seismic facies volume is a qualitative sand indicator for the M5. The yellow and red seismic facies generally align with sands in the M5 lobe (Figure 9). There are also differences between the two methods. SOM shows sand lobes with larger areas and suggests that sands are mostly absent in the eastern side of the survey. Inversion shows improved resolution and less contrast between the lobe and background deposition. The SOM suggests that sand is widely distributed across the area but inversion results show that the highest porosity sands are restricted to the northern sand lobes (Figure 8).

The northernmost sand lobe is approximately 26 mi² (67 km²) with an average net sand thickness of 81 ft (25 m) and average porosity of 26%. Average CO₂ density at storage conditions is 0.8 g/cm³, which is calculated from the average target depth of 7500 ft (2.3 km), normal pressure gradient, and a geothermal gradient of nearly 1F/100 ft. Variables such as pressure, temperature, fluid compressibility, salinity, porosity, and permeability can all impact the injection efficiency (⁠$Esaline$⁠). In addition to reservoir and fluid complexities, the overall development plan and execution can affect injection rates as well. This study is an initial, high-level estimate using efficiency factors calculated from a large distribution of site-specific projects in clastic reservoirs where P₁₀ = 4.6%, P₅₀ = 6.1%, and P₉₀ = 14.9% (Goodman et al., 2011). There is significant uncertainty around the inputs in this analysis. However, these initial parameters show the sands contain corresponding storage estimates of 16, 24, and 52 megatons (P₁₀, P₅₀, and P₉₀), respectively.

Key uncertainties for this analysis are compartmentalization and variations in the porosity and acoustic impedance trend. SOM and inversion outputs primarily impacted the area and porosity estimation for the final storage calculation. The gamma-ray logs show significant amounts of variation in sands and silts across the northern lobe area (Figure 9). These lower permeable zones could act as baffles and limit connectivity in the reservoir, thus increasing injection pressures during sequestration. Predictions of porosity also may require future updates as more data becomes available. The blind well result in Figure 7a shows that the inversion is overpredicting the measured impedance values, which would result in lower porosity values and underestimate the storage capacity. This deviation away from the strong tie observed in the control wells (Figure 5) could be an artifact due to the well’s proximity to a significant salt dome. Using a range of efficiency factors helps to capture these risks but more wells and additional log data would reduce the uncertainty.

These results show how 3D seismic and well data may be effectively used to characterize a potential CO₂ sequestration target. The demonstrated workflow is repeatable and can be replicated as researchers begin to characterize project-sized CO₂ storage targets:

1. Identify seismic attributes that identify geologic features for the given depositional/structural environment.

2. Run SOM with different subsets of seismic attributes until the desired features are outlined. In this study, SOM results provide a sand indicator and highlight deltaic sand lobes as potential carbon storage targets.

3. Invert for desired rock property (acoustic impedance) with a scope restricted to zones and formations highlighted by SOM.

4. Refine interpretations based on SOM and inversion to estimate geologic storage available. Inversion results restricted carbon storage estimates to only the sand lobe with best delineation and highest quality sand.

Results from SOM and AI elucidate the slope elements of a prograding fluvial deltaic system and highlight geologic features with better porosities for CO₂ storage. The study shows a workflow where SOM and acoustic impedance inversion maps are used in combination to characterize a deltaic lobe for possible carbon storage. The feature contains average net sands of 81 ft (25 m) and an estimated CO₂ storage capacity of 16, 24, and 52 megatons (P₁₀, P₅₀, P₉₀), respectively. This workflow can be applied in other areas in the southeastern United States where carbon storage projects are actively pursued. Future work is needed to address the risks from faults cutting across the reservoir, which present significant concern for long-term containment. Leakage also is possible from nearby wells penetrating the interval and would need further analysis.

We appreciate ConocoPhillips and Seismic Exchange Inc. for making seismic data available for use in this research. Thank you to TGS for giving approval for the well data displays. Seismic data are owned or controlled by Seismic Exchange Inc.; interpretations are those of the authors.

Data associated with this research are confidential and cannot be released.