AI-driven optimization of CO2 efficiency in CCUS projects

2.1. Field selection and characteristics

The selected fields span the Permian Basin and other major CO₂ EOR hubs, encompassing a range of heterogeneity, operational practices, and injection histories:

1) Denver Unit (DU): The Denver Unit in the Wasson Field is a major San Andres carbonate reservoir under CO₂ EOR since 1983, following pilot testing in 1978. It features strong surveillance programs and variable WAG ratios. The field is currently operated by Occidental Petroleum ‎[1, 13,‎ 14].

2) SACROC: The SACROC Unit in the Midland Basin began CO₂ injection in 1972, transitioning from immiscible to miscible floods over time. It targets the Pennsylvanian Canyon Reef with decades of EOR experience. Kinder Morgan is the current operator ‎[2, ‎17].

3) Wasson San Andres Field (WSA): The WSA represents a combination of units (e.g., Willard, Bennet Ranch, DU) in the San Andres formation. Discovered in 1936, it includes thick carbonate with extensive WAG-based EOR development since the 1980’s ‎[13].

4) Seminole San Andres Unit (SSAU): SSAU includes complex carbonate facies and a well-developed Residual Oil Zone. It has a long CO₂ EOR history with consistently high WAG ratios and known field-wide geological heterogeneity ‎[15].

5) East Vacuum (EV): East Vacuum is a smaller, data-rich field with phased CO₂injection. It serves as a valuable site for short-term CO₂ performance evaluation ‎[10-‎12].

6) Rangely Weber Sand (RWS): RWS lies in the Paradox Basin and is geologically distinct with a siliciclastic Weber Sand reservoir. Its behavior under CO₂ EOR differs from carbonate analogs, making it a valuable contrast field ‎[18,‎ 19].

2.2. Data acquisition and digitization

Historical production and injection data were digitized using Automeris.io ‎[17, ‎20], an open-source graphical digitization tool, from publicly available technical sources and additional literature and field summaries from DOE reports and technical presentations were also referenced to contextualize the data and validate trends ‎[1, 2,‎ 10-‎19].

Where necessary, data were smoothed using a moving average filter to minimize digitization noise. Missing values or inconsistencies were handled through interpolation or masking during model training. All-time series were aligned to a common index (month since CO₂ start), allowing for a standardized forecasting horizon across fields. See examples of digitized DU data ‎[1,‎ 13‎, 14] in Fig. 1, while EV field production data, shown in Fig. 2, was a combination of digitized data and publicly sourced data ‎[10‎-‎12].

Fig. 1Denver unit production and injection plot based on data digitized from public sources ‎[1‎, 13‎, 14]

Fig. 2EV production and injection plot based on data digitized from public sources ‎[10‎‎-12]

2.3. Parsing incremental CO₂ response and production

To isolate the oil production response attributable to CO₂ injection, the waterflood performance prior to CO₂ start was extrapolated using exponential decline curve analysis, following a common analytical practice for baseline estimation ‎[21]. The deviation from this trend during the CO₂ injection phase was attributed to incremental recovery from CO₂. This method is particularly important in mature fields where the transition from waterflood to CO₂ injection does not always exhibit a clean production breakpoint.

For SACROC, the start of the analysis window was aligned with the commencement of miscible CO₂ injection to exclude prior immiscible or waterflood effects. The parsed EOR response was used as the target variable in the TFT forecasting model, ensuring that predictions represent CO₂-specific incremental oil rather than total field production.

To isolate CO₂ production response, the gas-oil-ratio (GOR) prior to CO₂ injection was taken as a baseline and held constant throughout the CO₂ production period to extrapolate hydrocarbon gas production which enabled parsing out CO₂ production.

2.4. Derived metrics

To assess the performance of different WAG scenarios, several key operational metrics were computed:

1) CO₂ Utilization (MCF/BBL): Total CO₂ injected divided by incremental oil produced.

2) CO₂ Retention (%): Defined as 100 × (CO₂ injected − CO₂ produced) / CO₂ injected.

3) Gas-Water Ratio (GWR): Defined as MCFPD / BWIPD, used to characterize WAG regime.

4) Throughput: Calculated by converting volumetric injection rates using assumed formation volume factors.

5) Formation volume factors were assumed as follows:

– CO₂ FVF = 0.5, based on typical values reported for reservoir conditions in Permian fields ‎[14].

– Water FVF = 1.0, assuming negligible compressibility for brine under reservoir pressure.

These parameters enabled the computation of an efficiency score for each scenario, defined as the ratio of retention to utilization. This score provides a quantitative basis for selecting optimal WAG ratios that balance CO₂ storage goals with oil recovery efficiency.

3.1. Forecasting model: temporal fusion transformer (TFT)

The Temporal Fusion Transformer (TFT) is a state-of-the-art deep learning model for interpretable multi-horizon time series forecasting. Originally proposed by Lim et al. (2021), TFT combines long- and short-term memory components through LSTM layers with attention mechanisms, allowing it to capture both temporal dependencies and variable importance across time and feature dimensions.

TFT is particularly well-suited for CO₂ EOR forecasting due to its ability to incorporate static, time-varying known, and unknown covariates, its support for missing values and irregular time steps, and its scalability across multiple fields with differing temporal lengths.

Model implementation was done using the pytorch-forecasting library in Python with pytorch-lightning for training orchestration. All training was conducted on CPU to ensure reproducibility across environments.

3.2. Data preprocessing

For each of the six fields, time series data was standardized using the following preprocessing steps:

1) Time Re-indexing: All monthly data were aligned using a continuous time_idx variable starting from the onset of CO₂ injection (or miscible injection in SACROC).

2) Scaling: The target variable (incremental EOR oil) was normalized using a StandardScaler, and scaled values were used during model training.

3) Covariate Encoding: Time-varying known variables included:

– BWIPD (barrels of water injected per day).

– MCFID (thousand cubic feet of CO₂ injected per day).

– Month number.

These covariates informed the model of known injection conditions at each timestep. Each field was treated independently, and group normalization was applied to the target variable.

3.3. Parsing CO₂ response

The total oil production includes contributions from both waterflood and CO₂ injection. To isolate the incremental CO₂ response, an exponential decline model was fit to the pre-CO₂ waterflood period and extrapolated into the CO₂ injection period. The difference between the extrapolated curve and actual production was taken as the net CO₂-driven response. Figs. 3-4 show an example of parsing out the incremental oil from DU and EV fields, respectively, using Exponential decline. It should be noted that EV field did not have as much data to fit the decline curve.

Fig. 3Example of EOR production extrapolation using decline curve analysis for DU field

Fig. 4EOR production extrapolation using decline curve analysis for EV field

This approach is consistent with analytical best practices in field studies and ensures that the model focuses on forecasting CO₂-responsive oil rather than general field behavior. This parsed CO₂ EOR response became the model’s forecast target (target_scaled).

3.4. Hyperparameter optimization

To improve the performance of the Temporal Fusion Transformer (TFT) model, hyperparameter optimization was carried out to tune key parameters including max_encoder_length, hidden_size, attention_head_size, and learning_rate. A genetic algorithm-based search was used to explore the parameter space and minimize the validation root mean squared error (RMSE). This automated tuning process ensured better generalization and forecasting accuracy across all six CO₂ EOR fields. The selected hyperparameters were applied consistently across scenarios to preserve comparability of results.

3.5. Model training and validation

Each field's data was split into:

1) Training set (80 %).

2) Validation set (10 %).

3) Test set (10 %).

A “TimeSeriesDataSet” object was created per field using:

1) max_encoder_length = 6 months.

2) max_prediction_length = 1 month (rolled iteratively for a 12-month horizon).

3) Training used the following configuration:

– Learning rate: 0.001.

– Hidden size: 64.

– Attention head size: 1.

– Dropout: 0.

– Loss function: RMSE.

– Early stopping: Patience of 200 epochs.

Performance was evaluated using RMSE, MAE, and $R^2$. Results showed strong training accuracy, particularly for EOR oil production, with average $R^2$ values exceeding 0.87 across fields. For example, Figs. 5-6 show the comparison between actual and model predicted incremental oil production and CO₂ production, respectively for DU field, while Figs. 7-8 show the oil and CO₂ production for DU, respectively.

Fig. 5Actual vs predicted incremental oil production for DU

Fig. 6Actual vs predicted CO2 production for DU

Fig. 7Actual vs predicted incremental oil production for EV

Fig. 8Actual vs predicted CO2 production for EV

3.6. WAG scenario design and forecasting

To simulate the effect of different gas-water ratios (GWR) in WAG operations, five scenarios were developed per field using the GWR multipliers of $X=$ 1 (base), 2, 3, 0.5, 0.33.

For each scenario, water and CO₂ injection rates were recomputed to maintain a constant injection volume and throughput while altering the gas/water ratio:

where $C$ and $W$ are the original CO₂ and water injection rates, respectively. These modified rates were then input into the TFT model as future known covariates. Each scenario was forecasted for 12 months into the future, generating predictions for both CO₂ production (MCFPD) and EOR oil production (BOPD).

Separate models were trained for each target. For each forecast, the predicted oil and gas responses were used to compute:

– Utilization (MCF injected per BBL of oil).

– Retention (%).

– Efficiency score = Retention / Utilization.

These metrics allowed quantitative ranking of WAG ratios.

4.1. Model forecast accuracy

The Temporal Fusion Transformer demonstrated high accuracy in forecasting both CO₂ and oil production across all six fields. Forecast quality was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination ($R^2$).

The model exhibited consistently strong performance in oil prediction, with $R^2$ values exceeding 0.87 in most fields and peaking at 0.95 in the Wasson San Andres (WSA) field. CO₂ production forecasts also showed excellent alignment, achieving $R^2$ values above 0.91 in five out of six fields, except for East Vacuum (EV), which recorded lower accuracy ($R^2=$0.54), likely due to greater operational noise and limited waterflood production data to fit the exponential decline too. Summary of accuracy metrics for both Oil and CO₂ production are shown in Table 1 and Table 2, respectively.

These results validate the applicability of TFT for short-term forecasting of CO₂ EOR responses, providing a robust alternative to full-physics simulation for annual planning.

4.2. Optimization of WAG ratios

To evaluate the operational impact of varying WAG ratios, five injection scenarios were simulated per field, adjusting the CO₂-to-water injection ratio (GWR) while holding total injection volume approximately constant. Key metrics defined earlier included utilization, retention and efficiency score. Fig. 9 shows the efficiency trends for all fields with the colors corresponding to the GWR multiplier “X”. It can be seen from the figure that there is room to optimize most of these fields since the base case ($X=$1) does not always exhibit the highest efficiency score. This data is also summarized in Table 3. A general pattern emerged: GWR = 2-3 produced the highest retention (up to 79 %) but diminishing oil gains, leading to moderate efficiency scores. GWR = 0.3-0.5 offered better oil rates but negative retention due continuation of production of pre-injected CO₂. An intermediate GWR of ~1.0-1.5 often balanced both goals, achieving retention of 50-60 % with utilization rates between 13-21 MCF/BBL.

Table 4 presents the recommended adjustments to the gas-water ratio (GWR) for each field based on short-term Temporal Fusion Transformer (TFT) forecasts. The goal was to maximize CO₂ sequestration efficiency, defined here as the ratio of CO₂ retention to utilization (MCF injected per barrel of oil produced). For most fields, adjusting the GWR yields measurable improvements in efficiency. In the Denver Unit (DU), increasing the GWR from 0.9 to 1.7 results in an 8 % gain in efficiency, with CO₂ utilization improving by 38 % and retention by 48 %. SACROC shows a modest efficiency gain (+14 %) when increasing the GWR to 1.9, driven by significant improvements in retention (+96 %). Conversely, fields such as SSAU and EV benefit from reducing the GWR. In SSAU, decreasing the GWR from 0.9 to 0.5 leads to the largest relative improvement in efficiency (+31 %), despite reductions in both utilization and retention. For EV, reducing the GWR to 0.2 increases efficiency by 11 %, though model uncertainty (low R²) warrants cautious interpretation. Fields like RWS show minimal sensitivity to GWR changes within the tested range, suggesting that other operational levers may be more impactful. Meanwhile, WSA shows that pushing to higher GWR values (3.2) can yield a 19 % efficiency improvement, albeit with a notable increase in CO₂ injection rate.

Fig. 9Efficiency scores for different fields at different WAG multiplier scenarios

These findings support field-specific, short-term WAG adjustments as a practical lever for improving CO₂ management performance in mature EOR assets.

4.3. Operational insights

From the modeling exercise, several operational insights emerged. Short-term optimizations (12-month horizon) can be meaningfully supported with machine learning tools that are data-driven and less dependent on static reservoir models. Additionally, the efficiency score proposed here enables simultaneous targeting of both climate (sequestration) and economic (oil recovery) KPIs. This study shows that deep learning forecasting tools, particularly TFT, can play a critical role in bridging production optimization and carbon management. The efficiency score developed here – linking CO₂ retention and utilization – serves as a proxy for carbon abatement effectiveness. By embedding such models into operational workflows, operators can make evidence-based adjustments to WAG ratios, directly enhancing sequestration permanence while maintaining oil revenue streams. These insights are especially useful for compliance-driven projects targeting carbon credit monetization under programs such as 45Q and LCFS.