The SPE has split the former "Management & Information" technical discipline into two new technical discplines:
- Data Science & Engineering Analytics
The SPE has split the former "Management & Information" technical discipline into two new technical discplines:
|Theme||Visible||Selectable||Appearance||Zoom Range (now: 0)|
Razak, Syamil Mohd (University of Southern California) | Cornelio, Jodel (University of Southern California) | Cho, Young (University of Southern California) | Liu, Hui-Hai (Aramco Americas) | Vaidya, Ravimadhav (Aramco Americas) | Jafarpour, Behnam (University of Southern California)
Abstract Data-driven methods have surged in popularity due to increased field development and data collection effort in the last two decades, and partly because flow physics in hydraulically fractured low-permeability formations is poorly understood. Such statistical tools have limited extrapolation ability and require sufficient training data, where training an under-determined neural network predictive model with limited data can result in overfitting and poor prediction performance. Unlike statistical models, physics-based models impose causal relations that can provide reliable predictions over a wide range of input. While a detailed physics-based description of fluid flow in unconventional reservoirs is not yet available, approximate physical flow functions have been proposed to capture the general production behavior of unconventional wells. These physical functions can be augmented with the available data to enhance the extrapolation power of data-driven methods and constrain the output to adhere to the general production trends. We demonstrate two physics-constrained approaches (i.e., statistical approach and explicit approach) where physical flow functions are embedded into neural network models. The performance of physics-constrained models is however dependent on the relevance of the embedded physics to the observed data. When the data cannot be fully represented by a physics-constrained model, the resulting prediction for any given input comes with a large residual error when compared to the ground- truth. We further employ residual learning and introduce an auxiliary neural network component to learn the complex relationship between the input parameters (such as formation and completion properties) and the expected residuals that represent the imperfect descriptions or uncaptured physical phenomena. In this paper, the integration of residual learning and physics-constrained models constitute the Physics- Guided Deep Learning (PGDL) model. The PGDL model augments the predictions from the residual learning model and the physics-constrained model resulting in final predictions that significantly reduce the amount of under and over estimations for a more robust production prediction. Several synthetic datasets with increasing complexity as well as a field dataset from Bakken are used to demonstrate the performance of the proposed PGDL model.
Al-Jarro, Ahmed (Saudi Aramco) | Alshehri, Ali (Saudi Aramco) | Amer, Ayman (Saudi Aramco) | Althobaiti, Abdulrahman (Saudi Aramco) | Saiari, Hamad (Saudi Aramco) | Alsalman, Fatima (Baker Hughes) | Albaqshi, Muntathir (Baker Hughes) | Chan, Godine (Baker Hughes) | Kakpovbia, Anthony (Baker Hughes) | Kairiukstis, Edvardas (Baker Hughes) | Odisio, Matthias (Baker Hughes) | Qian, Weiwei (Baker Hughes) | Roy, Arjun (Baker Hughes) | Saad, Bilal (Baker Hughes) | Shapiro, Vladimir (Baker Hughes)
Abstract In this work, we present a novel Artificial Intelligence (AI) powered non-destructive testing (NDT) system for the detection of potential corrosion under insulation (CUI) inspections, code named DPCUI, developed by the Research and Development Center (R&DC) at Saudi Aramco in partnership with Baker Hughes. This inspection system enables fast external thermographic screening of large facilities by covering many condition monitoring locations (CML), and without any contact with the asset surfaces. It examines temporal thermography datasets that are collected using a high-resolution IR camera, such as those provided by FLIR, and on one or more RGB images that provide context of inspected areas. The collected data is analyzed by a dedicated AI engine to detect the presence of abnormal heat transfer signatures that occur due to defects present within the targeted CMLs. The novelty of this AI powered technology has several advantages. It provides a contact-less, smart, easy, fast, automated, safe, and reliable risk-based repair and maintenance decision making on the integrity of assets, enabling asset owners to efficiently prioritize their operations and processes in a seamless manner while the assets are kept online. Here, we enhance and extend the performance of our AI models to predict not only the presence of potential CUI or not, i.e. binary classification, but also various types of potential CUI, i.e. multi-class classification, a first of its kind.
Mohd Razak, Syamil (University of Southern California) | Cornelio, Jodel (University of Southern California) | Cho, Young (University of Southern California) | Liu, Hui-Hai (Aramco Americas) | Vaidya, Ravimadhav (Aramco Americas) | Jafarpour, Behnam (University of Southern California)
Abstract Neural network predictive models are popular for production forecasting in unconventional reservoirs. They have the ability to learn complex input-output mapping between well properties and observed production responses from the large amount of data collected in the field. Additionally, the flow behavior in hydraulically fractured unconventional reservoirs is not well understood making such statistical models practical. Variants of neural networks have been proposed for production prediction in unconventional reservoirs, offering predictive capability of varying levels of granularity, accuracy and robustness against noisy and incomplete data. Neural network predictive models that incorporate physical understanding are especially useful for subsurface systems as they provide physically sound predictions. In this work, we propose a new Dynamic Physics-Guided Deep Learning (DPGDL) model that incorporates physical functions into neural networks and uses residual learning to compensate for the imperfect description of the physics. The new formulation allows for dynamic residual correction, avoids unintended bias due to less-than-ideal input data, and provides robust long-term predictions. The DPGDL model improves upon a static formulation by utilizing a masked loss function to enable learning from wells with varying production lengths and by improving the results when partially-observed timesteps are present. We also develop a new sequence-to-sequence residual model to correct additional biases in the long-term predictions from the physics-constrained neural networks. Several synthetic datasets with increasing complexity as well as a field dataset from Bakken are used to demonstrate the performance of the new DPGDL model.
Tariq, Zeeshan (King Abdullah University of Science and Technology) | Xu, Zhen (King Abdullah University of Science and Technology) | Gudala, Manojkumar (King Abdullah University of Science and Technology) | Yan, Bicheng (King Abdullah University of Science and Technology) | Sun, Shuyu (King Abdullah University of Science and Technology)
Abstract Naturally fractured reservoirs (NFRs), such as fractured carbonate reservoirs, are ubiquitous across the worldwide and are potentially very good source to store carbondioxide (CO2) for a longer period of time. The simulation models are great tool to assess the potential and understanding the physics behind CO2-brine interaction in subsurface reservoirs. Simulating the behavior of fluid flow in NFR reservoirs during CO2 are computationally expensive because of the multiple reasons such as highly-fractured and heterogeneous nature of the rock, fast propagation of CO2 plume in the fracture network, and high capillary contrast between matrix and fractures. This paper presents a data-driven deep learning surrogate modeling approach that can accurately and efficiently capture the temporal-spatial dynamics of CO2 saturation plumes during injection and post-injection monitoring periods of Geological Carbon Sequestration (GCS) operations in NFRs. We have built a physics-based numerical simulation model to simulate the process of CO2 injection in a naturally fractured deep saline aquifers. A standalone package was developed to couple the discrete fracture network in a fully compositional numerical simulation model. Then reservoir model was sampled using the Latin-Hypercube approach to account for a wide range of petrophysical, geological, reservoir, and operational parameters. The simulation model parameters were obtained from extensive geological surveys published in literature. These samples generated a massive physics-informed database (about 900 simulations) that provides sufficient training dataset for the Deep Learning surrogate models. Average Absolute Percentage Error (AAPE) and coefficient of determination (R) were used as error metrics to evaluate the performance of the surrogate models. The developed workflow showed superior performance by giving AAPE less than 5% and R more than 0.95 between ground truth and predictions of the state variables. The proposed Deep Learning framework provides an innovative approach to track CO2 plume in a fractured carbonate reservoir and can be used as a quick assessment tool to evaluate the long term feasibility of CO2 movement in fractured carbonate medium.
Ma, Zheren (Quantum Reservoir Impact LLC) | Davani, Ehsan (Quantum Reservoir Impact LLC) | Ma, Xiaodan (Quantum Reservoir Impact LLC) | Lee, Hanna (Quantum Reservoir Impact LLC) | Arslan, Izzet (Quantum Reservoir Impact LLC) | Zhai, Xiang (Quantum Reservoir Impact LLC) | Darabi, Hamed (Quantum Reservoir Impact LLC) | Castineira, David (Quantum Reservoir Impact LLC)
Abstract Data-driven decisions powered by machine-learning methods are increasing in popularity when it comes to optimizing field development in unconventional reservoirs. However, since well performance is impacted by many factors (e.g., geological characteristics, completion design, well design, etc.), the challenge is uncovering trends from all the noise. By leveraging basin-level knowledge captured by big data sculpting, integrating private and public data with the use of uncertainty quantification, Augmented AI (a combination of expert-based knowledge and advanced AI frameworks) can provide quick and science-based answers for well spacing and fracking optimization and assess the full potential of an asset in unconventional reservoirs. Augmented AI is artificial intelligence powered by engineering wisdom. The Augmented AI workflow starts with data sculpting, which includes information retrieval, data cleaning and standardization, and finally a smart, deep and systematic data QC. Feature engineering generates all the relevant parameters going into the machine learning model—over 50 features have been generated for this work and categorized. The final step is to perform model tuning and ensemble, evaluating the model robustness, generating model explanation and uncertainty quantification. Augmented AI adopts an iterative machine learning modeling approach. This approach combines new and innovative engineering and G&G workflows with data-driven models so that a deep understanding of the field behavior can be developed. Loops from feature selection to model tuning are used until good model results are achieved. The loop is automated using Bayesians optimization. All machine learning models have different strengths and weaknesses for prediction. Instead of manually determining which machine learning model to use, this approach uses an adaptive ensemble machine learning approach that is a stacking algorithm that combines multiple regression models via a second level machine learning model. It smartly aggregates opinions from different models with reduced variance and better robustness. Augmented AI has been applied in unconventional reservoirs with great results. A case study in Midland Basin is presented in this paper. Domain-induced feature engineering was performed to obtain important features for predicting well performance, and initial feature selection was conducted using feature correlation analysis. A trusted and explainable ML model was built and enhanced with uncertainty quantification. After running several sensitivity analyses, Augmented AI optimized the attributes of interest, then vetted the outcome, generating a report and visualizing the results. In addition, further information about the direct impact of well spacing on EUR was deconvoluted from other parameters using an ML explanation technique for Wolfcamp Formation in Permian Basin and subsequently well spacing optimization was presented for the case study in Midland Basin. An innovative model was created using Augmented AI to optimize well spacing, leveraging big data sculpting, domain and physics-induced feature engineering, and machine learning. The learning was transferred from the basin model to the specific region of interest. Augmented AI provides efficient and systematic private data organization, an explainable machine learning model, robust production forecast with quantified uncertainty and well spacing and frac parameters optimization. Augmented AI models are already built for major basins such as Midland and Delaware basins. The learning and knowledge of the model can be transferred to any region in a basin and can be refined using more accurate private data. This allows conclusions to be drawn even with a limited number of wells.