Accurate and rapid forecasting of CO2 trapping, mobility, and plume evolution under dynamic injection conditions is crucial for effective planning, operational optimization, and regulatory compliance in geological carbon storage (GCS) projects. Traditional analytical methods often rely on oversimplified assumptions, compromising accuracy, while numerical simulations, though precise, require extensive computational resources, limiting their utility for real-time scenario analysis.To overcome these challenges, this study proposes an advanced deep learning framework for forecasting trapped and movable CO2 fractions and plume extent. An enhanced sequence-to-sequence (Seq2Seq) neural network with a composite loss function robustly predicts CO2 volumetrics, while a hybrid Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) model forecasts CO2 plume extent. The models incorporate nine static geological and reservoir characteristics and dynamic injection profiles, capturing real-world injection complexities, including varying rates and intermittent schedules with multiple start-stop events. Training utilized several hundred high-fidelity numerical simulation realizations covering diverse geological and operational scenarios.The enhanced Seq2Seq model achieved an average Mean Absolute Error (MAE) of 0.016 for trapped and movable CO2 fractions, while the LSTM-MLP model attained an average MAE of 42 meters for plume diameter. These deep learning-driven surrogates drastically reduce computational time, providing accurate forecasts within seconds per scenario compared to conventional methods requiring hours. This significant advancement facilitates rapid, reliable decision-making, optimized storage strategies, and rigorous regulatory compliance in GCS initiatives.

Reliable prediction of CO2 plume extent evolution is essential for ensuring safe and effective geological carbon storage. Accurate plume forecasts help prevent leakage, reduce environmental risks, and verify that the injected CO2 remains securely confined within the intended storage formation. In this study, a machine learning–based forecasting framework was developed to efficiently predict plume extent at multiple depths in a saline aquifer. A total of 2,000 distinct geological carbon storage scenarios were simulated using a numerical reservoir model. These scenarios were created by systematically varying key geological and operational parameters, including permeability, salinity, compressibility, reservoir thickness, pressure, and injection rates. Plume extent predictions were generated at six reservoir depths across 81 discrete time steps over an 80-year simulation period that included 30 years of continuous CO2 injection, representing a total of 30 million metric tons of CO2, followed by 50 years of post-injection monitoring. To enhance computational efficiency and forecasting accuracy, Singular Value Decomposition was incorporated into the prediction workflow. Four machine learning approaches were evaluated, namely Random Forest, Multilayer Perceptron, Long Short-Term Memory, and Gated Recurrent Unit, both with and without dimensionality reduction. The results demonstrated that the Multilayer Perceptron without dimensionality reduction achieved the highest accuracy, producing a mean absolute error of 23 meters. Forecasts of plume extents across all reservoir layers for a single scenario were obtained in less than one second, enabling rapid three-dimensional visualization of plume geometry. The developed framework accurately predicts plume extents of up to 7,200 meters in upper reservoir layers and up to 2,000 meters in deeper layers, while reliably capturing the dynamics of the CO2 saturation front. The principal contribution of this study is the development of a fast, accurate, and generalizable prediction workflow that can be applied across a wide range of geological and operational settings, thereby providing critical support for decision-making and risk management in geological carbon storage projects.

This study aims to enhance the accuracy and computational efficiency of pressure and saturation predictions in geological carbon storage (GCS) projects. By leveraging an improved Fourier Neural Operator (FNO) architecture and implementing transfer learning (TL), the research seeks to enable reliable CO2 storage forecasting under diverse geological and operational conditions. The study employs an advanced FNO model, optimized through modifications such as additional Fourier layers, convolutional layers, and improved padding schemes to enhance predictive performance. Transfer learning is integrated to adapt pre-trained FNO models for varying injection rates, intermittent injection schedules, and variable injection locations. The models are trained and validated using simulated data from the SACROC geological model, a heterogeneous saline aquifer, with testing across multiple operational scenarios. The improved FNO models significantly outperform baseline FNO architectures, reducing training time and data requirements by 98 % and 60 %, respectively. The mean absolute errors (MAE) for pressure predictions are 0.27 MPa (variable injection rates), 0.51 MPa (intermittent injection), and 0.59 MPa (variable injection locations). Saturation predictions exhibit MAE values of 0.05 (variable injection rates), 0.056 (intermittent injection), and 0.14 (variable injection locations), with reduced accuracy in the latter due to FNO's assumptions on fixed source points. The results demonstrate that transfer learning enables FNO models to generalize effectively across diverse geological scenarios while maintaining high predictive accuracy. This study advances the state of knowledge in GCS modeling by demonstrating the effectiveness of transfer learning in adapting FNO models to complex, real-world injection scenarios. The proposed improvements address key limitations in existing ML-based forecasting approaches, paving the way for more efficient, data-driven decision-making in subsurface CO2 storage operations. These findings contribute to the practical implementation of ML-based forecasting for regulatory compliance and operational optimization in GCS projects.

Underground hydrogen storage (UHS) must balance injectivity, productivity, hydrogen purity, withdrawal efficiency, and hydrogen-to-water ratio over repeated cycles and across geological conditions. We present a unified optimization framework that uses simulation-trained machine-learning (ML) proxies to co-optimize these five metrics with respect to injection/production rates, half-cycle duration, cushion-fill time, and cushion-gas composition. The proxies were trained on 500 high-fidelity compositional simulations generated from a synthetic aquifer reservoir model and achieved held-out test accuracies of R2 = 0.905-0.953 across the five metrics. These cycle-resolved models were embedded in a global differential-evolution search using a normalized multimetric area-under-curve objective. Unlike most prior studies, the framework simultaneously evaluates pure, binary, and ternary CO2-CH4-N2 cushion gases while treating cycle timing as a decision variable. Results show that operational controls dominate performance: the best composite outcomes occur at high throughput with relatively short half-cycles (1-3 months) and cushion-fill times typically in the 2-10 month range. Methane (CH4) is the most reliable baseline cushion gas across most cases, while a CO2-CH4-N2 ternary blend improves injectivity in fast-cycling regimes. Across reservoir-quality cases, the optimizer consistently retains CH4 and a 2-month half-cycle, while reducing cushion-fill time from 12 months (lower-quality case) to 4 months (medium/highquality cases). The workflow transfers to a heterogeneous 3D geomodel for rapid site-specific screening.

Geological model compression is crucial for making large and complex models more manageable. By reducing the size of these models, compression techniques enable efficient storage, enhance computational efficiency, making it feasible to perform complex simulations and analyses in a shorter time. This is particularly important in applications such as reservoir management, groundwater hydrology, and geological carbon storage, where large geomodels with millions of grid cells are common. This study presents a comprehensive overview of previous work on geomodel compression and introduces several autoencoder-based deep-learning architectures for low-dimensional representation of modified Brugge-field geomodels. The compression and reconstruction efficiencies of autoencoders (AE), variational autoencoders (VAE), vector-quantized variational autoencoders (VQ-VAE), and vector-quantized variational autoencoders2 (VQ-VAE2) were tested and compared to the traditional singular value decomposition (SVD) method. Results show that the deep-learning-based approaches significantly outperform SVD, achieving higher compression ratios while maintaining or even exceeding the reconstruction quality. Notably, VQ-VAE2 achieves the highest compression ratio of 667:1 with a structural similarity index metric (SSIM) of 0.92, far surpassing the 10:1 compression ratio of SVD with a SSIM of 0.9. The result of this work shows that, unlike traditional approaches, which often rely on linear transformations and can struggle to capture complex, non-linear relationships within geological data, VQ-VAE's use of vector quantization helps in preserving high-resolution details and enhances the model's ability to generalize across varying geological complexities.

The rapid advancement of machine learning techniques, particularly Fourier Neural Operators (FNO), offers a promising approach to predicting CO2 saturation and pressure distributions in geological carbon storage. This study explores the application of FNO combined with transfer learning (FNO+TL) to enhance computational efficiency and accuracy in forecasting CO2 storage under diverse geological and operational conditions. We trained FNO models on datasets from the SACROC (153 samples) geological model and applied TL to predict outcomes for the Illinois Basin-Decatur Project (IBDP). Our findings highlight the substantial computational savings without significant compromise in performance of the FNO+TL models compared to FNO, using 10 and 20 samples for pressure and saturation predictions respectively. The FNO+TL model achieved an average Mean Absolute Error (MAE) of 0.11 for CO2 saturation and 8.7 psia for pressure predictions, compared to 0.79 and 2.4 psia respectively for FNO. While saturation predictions were less precise, the model effectively captured the overall CO2 migration trends. Notably, transfer learning significantly reduced computational costs, decreasing training time by 62.5 % and storage, RAM requirements by 90 % and 68 %, respectively. Despite some limitations in saturation prediction accuracy, the FNO+TL approach demonstrates potential for efficient and reliable CO2 storage forecasting. This study highlights the potential of FNOs and transfer learning for efficient and accurate forecasting of CO2 storage behavior and management of carbon sequestration projects.

Geological carbon storage (GCS) is critical for sequestering CO2 deep underground. GCS projects may face environmental challenges, such as leakage risks, adverse pressure buildup, and groundwater contamination. Numerical simulators play a vital role in accurate forecasting but can be computationally expensive. In this work, we leveraged an updated Fourier Neural Operator (FNO) which includes data sparsity management, to learn to rapidly forecast pressure and CO2 phase saturation distributions in a geological carbon storage (GCS) reservoir. Compared to commercial reservoir simulators, FNO-based forecasting offers accurate prediction while reducing the computational time by a factor of 40, enabling high volume of forecasting in less time. Additionally, we applied transfer learning (TL) to further reduce the data and computational requirements of the FNO-based forecasting across a wide array of scenarios. Specifically, we demonstrated the usefulness of TL in accurately predicting the pressure and CO2 saturation distributions for uncertain and variable geological and operational conditions. The results of this study indicate that the improved FNO workflow reduces the computational time by approximately 97 %, and the relative mean error for predicting both CO2 saturation and pressure distributions is <1 %. Generally, the use of TL effectively transfers knowledge from a pre-existing model to other related tasks. TL significantly reduces the required training data by 78 % while maintaining a relative mean error below 5 %. Although, the results in this work can be further improved, this study demonstrates the potential of integrating FNO and TL to reduce computational time and data requirements for CO2 forecasts during GCS projects, providing a more efficient and faster approach.

Abstract Correlations between the fluid injection via water disposal (WD) wells and hydraulically fractured (HF) wells into subsurface earth and the subsequent increase in seismic activity is well documented. Our research presents a method to go beyond statistical correlations and quantify the causal relationship between subsurface fluid injection and induced seismic events, accounting for the confounding factors. The dataset employed for this analysis covers a 7-year period within the state of Oklahoma. The dataset comprises details such as the locations and key operational metrics of the water disposal and hydraulically fractured wells. Additionally, the dataset contains the location and magnitude of earthquakes, as well as the location and length of major fault lines. For this study, earthquakes (which can be induced seismicity) have a magnitude higher than 2. Our dataset contained 22,368 earthquakes ranging from magnitude 2 to 5.8 with a mean of 2.47 and mode of 2.2. Utilizing double machine learning (DML), we estimate the average treatment effect (ATE) to precisely quantify the causality. ATE results were derived from diverse spatiotemporal sample combinations, pinpointing the areal extent and temporal duration of heightened causality. This enabled us to explore the relationships between seismic activity and WD/HF wells in terms of the areal and temporal extents of the effects of the fluid injection into the subsurface through WD or HF wells. The ATE results reveal that 2 active water-disposal wells over 56 days cause 1 earthquake within a 4,400 sq. km area. Hydraulically fractured wells were found to have a more localized causal impact, such that 3 hydraulically fractured wells over 106 days lead to 1 earthquake within a 200 sq. km area. No detectable causal effect of fluid injection on earthquakes magnitude was identified.

Abstract Subsurface earth models, also known as geomodels, are essential for characterizing and developing complex subsurface systems. Traditional geomodel generation methods, such as multiple-point statistics, can be time-consuming and computationally expensive. Generative Artificial Intelligence (AI) offers a promising alternative, with the potential to generate high-quality geomodels more quickly and efficiently. This paper proposes a deep-learning-based generative AI for geomodeling that comprises two deep learning models: a hierarchical vector-quantized variational autoencoder (VQ-VAE-2) and a PixelSNAIL autoregressive model. The VQ-VAE-2 learns to massively compress geomodels into a low-dimensional, discrete latent representation. The PixelSNAIL then learns the prior distribution of the latent codes. To generate a geomodel, the PixelSNAIL samples from the prior distribution of latent codes and the decoder of the VQ-VAE-2 converts the sampled latent code to a newly constructed geomodel. The PixelSNAIL can be used for unconditional or conditional geomodel generation. In unconditional generation, the generative workflow generates an ensemble of geomodels without any constraint. In conditional geomodel generation, the generative workflow generates an ensemble of geomodels similar to a user-defined source geomodel. This facilitates the control and manipulation of the generated geomodels. To improve the generation of fluvial channels in the geomodels, we use perceptual loss instead of the traditional mean absolute error loss in the VQ-VAE-2 model. At a specific compression ratio, the proposed Generative AI method generates multi-attribute geomodels of higher quality than single-attribute geomodels.

Rapid simulation of the spatiotemporal evolution of pressure & saturation for SACROC 1. Neural operator was trained on only 153 simulation runs 2. Trained to account for heterogeneity and variations/uncertainties in engineering, fluids, and geology 3. Pressure forecast has less than 1% error 4. Saturation forecast has less than 2% error 5. Traditional simulator takes 1 hour for a single scenario, while neural operator takes less than 1 minute. Rapid simulation of the spatiotemporal evolution of pressure & saturation for IBDP 1. Transfer Learning was implemented on the SACROC-based Neural Operator that was trained on only 20 simulation runs for IBDP Site 2. SACROC and IBDP Sites have several significant differences in geology and engineering parameters. 3. Pressure forecast has less than 5 psi error 4. Saturation forecast has less than 7% error 5. Traditional simulator takes 1 hour for a single scenario, while neural operator takes less than 1 minute and only 20 simulations for training/validation.

Abstract Production forecasting is vital in the oil and gas sector, empowering engineers with insights for effective reservoir management. This paper introduces the concept of Transfer Learning as a powerful tool in the domain of machine-learning-assisted production forecasting that accounts of 3D spatial distributions of three geological properties, namely porosity, permeability, and saturation, two completion parameters, namely hydraulic fracture height and length, and production constraint. Transfer learning efficiently leverages knowledge from one problem to improve generalization on another, especially when data is scarce and computational resources are limited. To demonstrate the utility of transfer learning, we evaluate two scenarios of transfer learning. The first transfer learning scenario demonstrates the generalization of the forecasting to cases with variable hydraulic fracture spacing using limited training data. The second transfer learning scenario demonstrates the generalization of the forecasting to cases with variable natural fracture spacing and natural fracture permeability using limited training data. Source dataset contained 2000 realizations, while the target dataset contained 20, 40, 60, 80, 100, 250, 500, or 1000 realizations to represent the scenarios of limited training-data availability. The study confirms the benefits of transfer learning when the training dataset size is small (generally less than 100 training realizations); however, under large training dataset size (around 500 or more training realizations), transfer learning is not needed. For the first scenario involving variable hydraulic fracture spacing, the use of transfer learning ensured that the source model can be trained on target dataset with 80 realizations for gas rate forecasting at an accuracy of 25% in terms of MAPE, and with 500 realizations for condensate rate forecasting at an accuracy of 12% in terms of MAPE. Similarly, for the second scenario involving variable natural fracture spacing and natural fracture permeability, the use of transfer learning ensured that the source model can be trained on target dataset with 250 realizations for gas rate forecasting at an accuracy of 24% in terms of MAPE, and with 500 realizations for condensate rate forecasting at an accuracy of 23% in terms of MAPE. This illustrates the potential of transfer learning in improving forecasting models with limited data using a well pre-trained model and enhanced hyperparameter tuning of the transfer learning model. For cases with 500 or more training realizations, transfer learning severely underperforms as compared to training a conventional machine-learning model from scratch. The paper explores two cases of transfer learning.

Abstract The study shows the use of unsupervised manifold learning on microseismic data for fracture monitoring and characterization. Manifold learning condenses complex, high-dimensional data into more concise, lower-dimensional representations that encapsulate valuable underlying patterns and structures of the data. The study leverages Uniform Manifold Approximation and Projection (UMAP) methodology to efficiently convert high-dimensional data into a graph-based, lower-dimensional representation. This transformative approach adeptly captures meaningful patterns and structures while preserving both local and global distances; thereby, the topology. In this study, the unsupervised manifold learning is applied on accelerometer and hydrophone data obtained during an intermediate field-scale hydraulic stimulation experiment conducted at the Sanford Underground Research Facility in South Dakota to measure, monitor, and characterize fracture propagation in near-real-time. Each micro-earthquake location identified using travel-time information is assigned a fracture type using unsupervised manifold learning, and later is assigned a fracture-plane label using semi-supervised manifold learning. Our findings highlight the precision and accuracy of our proposed method in preserving fracture network clusters and fracture-plane labels using signals from both accelerometers and hydrophones. These remarkable results underscore the potential of the Uniform Manifold Approximation and Projection (UMAP) technique as a versatile unsupervised-learning tool for unveiling the intrinsic structural features embedded within microseismic signals generated during hydraulic fracturing for purposes of fracture monitoring and characterization. Next, this study addresses the challenge of reliably assigning fracture-plane labels to microseismic events created during hydraulic stimulation. Most events lack clear labels due to scattering and uncertainty in event locations. To tackle this, the study introduces a semi-supervised learning strategy based on UMAP that utilizes a sparse dataset with pre-labeled fracture-plane categorizations. By combining event coordinates and Fourier spectra from geophone signals in a low-dimensional manifold, this approach achieves precision and recall rates exceeding 92% while using only about 20% of labeled events. This method remains robust even in the presence of location errors, making it valuable for enhancing fracture network characterization in various applications, including hydrocarbon and geothermal production.

In the upstream oil and gas industry, production forecasting is crucial for decision making, investment allocation, reservoir management, and field development. While traditional methods like empirical models and numerical simulators have been employed for production forecasting, their limitations become pronounced when dealing with large, heterogeneous reservoirs. Though traditional simulators provide detailed and accurate models of the reservoir, they can be computationally expensive and time consuming. Machine -learning based production forecasting can offer a faster yet accurate approach that is useful in scenarios where quick estimates are needed or thousands of scenarios need to be investigated within an optimization and history -matching workflow. We propose a machine learning driven workflow that integrates a massive geomodel compression followed by neural -network -based regression to rapidly forecast the production of a well in large, complex reservoir. The proposed workflow accounts for spatial heterogeneities in porosity, permeability, and fluid saturation as well as uncertainties induced and natural fracture properties and their distributions. The newly developed geomodel compression followed by neural -network -based rapid production forecasting reduces the computation time to generate a 5 -year forecast per realization by an order of 6. Overall, the newly developed rapid production forecasting per realization takes approximately 0.0003 s as compared to 1037 s or 17 min per realization for a traditional simulator.

This study investigates the spatial arrangement of existing crack pathways and the crack network at the time of crack coalescence in brittle rock-like materials, focusing on the factors that contribute to rapid and extensive crack growth immediately afterward. To that end, we extract relevant informative features from simulated crack networks generated by the HOSS simulator and employ machine learning methods to establish correlations between these features and the initiation of rapid and extensive crack propagation following coalescence. These features serve as indicators revealing the underlying mechanisms of crack propagation and coalescence. We utilize a Support Vector Machine (SVM) classifier with an RBF kernel to delineate the decision boundary between samples experiencing rapid and extensive crack propagation post-coalescence and those that do not. Through permutation feature importance evaluation, we identify the seven most crucial crack-network features associated with rapid and extensive crack growth. Notably, these features are particularly related to the status of energy buildup, energy distribution, and energy release inside the material. This study aims to elucidate the rapid and extensive fracture propagation post-coalescence in terms of the KDE-based and subgraph-based features. The fundamental assumption guiding this study is that direct observation of the stress and energy state within the material is not feasible. Consequently, understanding the phenomena of coalescence and extensive fracture propagation post-coalescence requires an exploration of the mechanisms observed through the spatiotemporal evolution of the crack network under uniaxial compression.

Abstract Deep learning can significantly accelerate the simulation of the injection, storage and production processes in an underground hydrogen storage (UHS). By understanding complex system interactions, deep learning offers accelerated simulations, enabling the creation of an fast visualization, forecasting and optimization framework. Our research presents a novel data-driven approach leveraging deep learning to mitigate the computational challenges of high-fidelity underground hydrogen-storage simulations. In this study, an innovative Fourier-Integrated Hybrid Neural Framework (F-IHNF) is used to create deep-learning-based surrogate models for field-scale hydrogen storage simulation. This framework combines Convolutional LSTM, 3D convolutions, and Fourier Neural Operators (FNO) for precise spatio-temporal analysis, focusing on hydrogen flow dynamics influenced by production and injection cycles. The deep-learning-based accelerated simulation workflow is developed and deployed on Fenton Creek gas reservoir model, segmented into 97×18×35 grid blocks with 61,110 active cells and a grid size of 121ft×136ft×2.8ft. The newly developed 3D Fourier-Integrated Hybrid Neural Framework (F-IHNF) achieves an impressive 98% accuracy in blind test validation for both pressure and hydrogen saturation forecasting over a period of 2 years involving 3 production-injection cycles. In the data generation phase, 76 simulations of hydrogen storage in various realizations of Fenton Creek geomodel under various engineering parameters were generated over 48 hours, with each 24-month hydrogen-storage simulation averaging 38 minutes. In comparison, the F-IHNF takes 0.5 seconds to forecast the spatiotemporal pressure and saturation evolutions over 24 months. This marks a 5000-time speedup in the forecasting. It took approximately 2 hours to train the F-IHNF model using 60 training samples and 6 validation samples. By finely tuning key hyperparameters—including convolutional LSTM hidden channels, 3D convolutional channels, and the selection of Fourier modes in x, y, and z directions—the precision of F-IHNF was impoved. The use of a weighted mean absolute error for saturation and a normal mean absolute error for pressure resulted in validation MAEs of 0.002 and 0.004, respectively. This study integrates a Fourier-Integrated Hybrid Neural Framework with deep learning for accelerated simulation of underground hydrogen storage. This approach will support faster analysis, visualization, forecasting, monitoring, and robust optimization, potentially aiding the adoption of underground hydrogen storage in the renewable energy sector for a sustainable future.