EAGE Conference on Energy Excellence: Digital Twins and Predictive Analytics

Fields in the Gulf of Thailand are currently focused on optimizing infill well placement to maximize extraction from both shared and unshared hydrocarbon reservoirs. This strategy utilizes an advanced Gulf of Thailand model that integrates specialized concepts of infill well spacing. A key objective is to identify new reservoirs to unlock additional potential from these wells.

Traditionally, reservoir connectivity is identified using wireline logging, pressure tests, and geological data. Relationships between the width and thickness of sandstone bodies in specific depositional environments help predict reservoir connectivity. The methodology involves manual identification of reservoir connectivity and facies based on well data. The distance to connected reservoirs is measured and projected to determine true channel width trends, validating width-to-thickness ratios by facies and units. A Random Forest regression model assesses parameter influences to predict true channel width and sand connectivity among wells.

The width-to-thickness ratios differ significantly between the channel (1:20 to 1:700) and bar (1:200 to 1:450) sandstone facies, reflecting distinct reservoir geometries and connectivity patterns. The model’s accuracy is validated using Leave-One-Out Cross-Validation, achieving high performance (80%-90%) across geological units. This predictive tool supports new reservoir identification and optimization of reservoir productivity in the Gulf of Thailand.

The study demonstrates the application of supervised machine learning algorithms to classify reservoir rock types using labeled data from well logs, cores, and other sources, improving accuracy and reducing manual interpretation time.

A 3D structural grid was constructed using Python, incorporating well tops, fault sticks, and depth horizons. Facies were predicted and populated throughout the reservoir zone using a Random Forest classifier algorithm.

Additional reservoir properties such as porosity, volume of shale, and saturation were populated across the entire grid using the XGBoost regressor algorithm, with the derived facies serving as an independent variable.

In-place resources were estimated using both deterministic and probabilistic methods, including sensitivity analysis, with estimates varying within 5% of those from a benchmark conventional model.

The study highlights that AI and ML methodologies can automate static modeling processes, facilitating data-centric reservoir description and swift updates to models. This is particularly beneficial for mature fields with extensive well data.

Identifying and quantifying infill targets are vital for optimizing ongoing oil field production. This study a Norwegian North Sea field by a novel digital approach that combines physics-based principles with data-driven techniques. This digital hybrid method generates historical fluid distribution maps, predicts future fluid movements, and identifies potential locations for new wells to enhance recovery. The performance of identified targets was predicted by advanced machine learning techniques. A comparison between the hybrid workflow results and results predicted by numerical simulation showed a high degree of consistency. This hybrid approach enables rapid, accurate digital field development planning, addressing a wide range of uncertainties and providing valuable insights for the E&P industry. The findings demonstrate that integrating data-driven and physics-based methods significantly enhances the ability to locate and quantify remaining oil, optimizing field development strategies.

This paper presents a novel approach for optimizing oil and gas field development using multi-level and proxy models integrated with machine learning algorithms. The traditional method of calculating Integrated Asset Models (IAMs) sequentially is time-consuming and often impractical for timely decision-making. Our proposed multi-level optimization approach accelerates IAM computations, enabling the evaluation of a greater number of development scenarios within shorter time frames. By utilizing machine learning to create fast, less detailed models, this method prioritizes scenarios based on predicted outcomes, ensuring a balance between computational efficiency and result accuracy. The approach also incorporates data from reservoir models, well models, and surface gathering and transportation systems to enhance model fidelity. This method not only reduces computation time but also facilitates rapid and accurate strategic decision-making, essential for maintaining competitiveness in the dynamic oil and gas industry. The implementation of this optimization strategy promises to improve the efficiency of field management and development, leading to higher extraction rates and increased profitability.

This study focuses on automating the creation of Amplitude Variation with Offset (AVO) sensitivity matrices using machine learning (ML) to improve seismic exploration. Traditional methods are labor-intensive and prone to inaccuracies due to manual processes and expert interpretation. We developed a Random Forest model to estimate Bulk Density (RHOB), Compressional Velocity (VP), and Shear Velocity (VS) in the Balingian area, Offshore Sarawak.

Our model integrated well log data and employed sequential prediction for enhanced accuracy. Data preprocessing involved removing outliers and bad data points. Key features included gamma ray, bulk density, and seismic velocities. Hyperparameter tuning via Grid Search with cross-validation ensured optimal performance.

Results showed the model performed well under simpler fluid conditions (INSITU and BRINE) but struggled with complex conditions (OIL and GAS), as indicated by higher error rates and negative R-squared (R²) values. Metrics such as Mean Absolute Percentage Error (MAPE) and Mean Absolute Error (MAE) highlighted these challenges.

This study demonstrates ML’s potential to automate and improve seismic velocity predictions, aiding petrophysics in subsurface understanding and drilling target identification. Further refinement and additional data are necessary to enhance predictive accuracy for complex fluid scenarios, paving the way for more efficient and accurate seismic exploration processes.

Conventional depofacies classification methods are often hindered by the massive data volumes, potential biases, and the labor-intensive nature of the process. Gamma ray logs are particularly effective for facies and sequence stratigraphy analysis due to their diverse shapes, higher resolution, and distinctive features. To enhance the accuracy and efficiency of depofacies prediction, this study employs a computer vision based deep learning approach to classify gamma ray logs into various categories of depofacies.

We applied convolutional neural networks (CNNs), known for their proficiency in handling graphical data. The volume of shale was derived from Gamma Ray (GR) by Steiber equation to identify the sand and shale interfaces, marking the top and bottom of these layers. Gamma ray (GR) logs were then preprocessed by converting them into images, cleansing the data, and dividing it into training and validation sets. The CNN model was trained to recognize coarsening upward, fining upward, bed, and serrated classes of gamma ray logs.

The model’s performance was assessed across a substantial number of images using a confusion matrix, achieving a training accuracy of 85%. This machine learning approach demonstrates significant potential in automating and improving the accuracy of well log classification which helps in identify depofacies.

Advancements in computing, data storage, and communication have revolutionized the oil and gas industry through artificial intelligence (AI), machine learning (ML), and big data. These technologies, while enhancing analysis efficiency, are complemented by numerical simulators to manage and optimize reservoirs. The integration of AI and ML with traditional methods, such as simulation models, may be used to improve forecast accuracy by addressing the limitations of each approach. Digital transformation is central to this evolution, automating previously manual processes, optimizing equipment operations, and improving decision-making through advanced data analysis and visualization tools. This study presents the Digital Transformation application to short-term management, control and forecast. Key components include real-time monitoring, predictive analytics and control, and a three-stage optimization process that enhance short-term and life-cycle reservoir performance. These technologies are implemented in UNISIM-IV Benchmark to test and validate the proposed workflow to improve reservoir short-term forecast, management and optimization. This comprehensive approach leads to increased productivity, cost reduction, and more sustainable practices.

This study introduces a novel generative deep learning-based methodology that enhances seismic resolution. This is crucial for optimizing field exploration and development as it enables more accurate geological interpretations. By integrating geological and geophysical knowledge with data-driven models, the methodology significantly improves subsurface imaging and inversion. Applied to the Chandon Field, offshore NW Australia, this approach has achieved substantial resolution enhancements, allowing for the detailed interpretation of previously unresolved stratigraphic and structural features. The improved seismic data and acoustic impedance models have enabled precise mapping of facies distribution and strati-structural plays, particularly within the Mungaroo and Brigadier Formations. The high-resolution seismic imaging has resulted in clearer visualization of structural and depositional features, facilitating confident fault interpretation and the identification of key geological structures, such as a strike-slip corridor. This work highlights the potential of deep learning combined with domain expertise to drive advancements in reservoir characterization and overall field management.

This study explores the development and application of digital twin technology to enhance field production and optimization through real-time decision support. It specifically focuses on optimizing production strategies and hydraulic fracturing job designs. The approach involves training an AI model with comprehensive field data, including reservoir properties, geological features, and operational parameters. By running numerous simulations using a multi-model dynamic approach and various AI algorithms, the model identifies optimal solutions, which are then validated through history matching, and identifying key performance indicators. Simulation outputs are integrated with financial modeling to reduce operational costs. The research highlights two key applications: optimizing production strategies and hydraulic fracturing job designs. For production strategy optimization, the digital twin model has led to a 6.3% increase in CBM field gas production over three months by fine-tuning parameters such as pump RPM and orifice bean size. In hydraulic fracturing, the model predicts optimal fracture length and permeability to enhance production efficiency, thereby identifying cost-effective fracking job design parameters. This AI-driven framework offers a powerful tool for real-time decision-making, significantly improving efficiency and profitability in field operations.

Reservoir fluid samples obtained from testing or production wells need to undergo PVT analysis to determine its properties under varying pressure, volume and temperature conditions. However, samples need to be tested for contamination, to ensure that the sample obtained is clean and representative of the downhole formation fluid, so that the PVT analysis gives reliable results.

During sampling, formation fluid is pumped through the wireline to the borehole. Initially, the fluid that flows will primarily consist of mud filtrate, and as pumping continues, the sample will gradually transition to a mixture of mud filtrate and formation fluid. Ideally, it is best for the sample to achieve the lowest possible contamination level, but it is impractical to pump indefinitely due to the cost of the rig time. The proposed solution, Live Integrated Sampling Advisory (LISA) leverages predictive modeling to perform real-time composition analysis of C1-C6+ hydrocarbon components to assess the contamination level of reservoir fluid. This could be implemented real-time, to aid reservoir engineers in determining the optimal time to sample reservoir fluid with contamination level within acceptance range.