Michael J. Pyrcz1, John Foster1,2, Carlos Torres-Verdín1, and Eric van Oort1
1 Hildebrand Department of Petroleum & Geosystems Engineering, the University of Texas at Austin
2 Institute for Computational Engineering and Science, the University of Texas at Austin
The participants in our kick off meeting requested that we provide a 'menu' / list of potential initial research projects. Note the actual consortium work is subject to member company steering. Research, prototype demonstrations and training materials included to support internal use of results.
- Training Models – machine learning requires a large set of training data. The member companies would benefit the construction of a suite of realistic numerical, consistent 3D multivariate earth models with structural and stratigraphic heterogeneity, petrophysical, geophysical and geomechanical properties, well and production data. We will account for realistic physics and messy, noisy, unbalanced, and incomplete measurements
- Fair Training and Testing Workflows – development of training and testing workflows that fairly assess the prediction accuracy of machine learning models. Note: current random selection of testing subsets is not fair. This includes improved methods for design of optimum model complexity and to diagnose model overfit.
- Data Preparation Methods – methods for normalization, standardization of subsurface features for input into machine learning. This includes data debiasing, imputation and accounting for measurement and interpretation uncertainty and spatial and scale context.
- Proxies for Reservoir Production – rapid production forecasting from subsurface modes of initial and enhanced production and recovery for real-time feedback of reservoir production from subsurface modeling decisions and also for inference of subsurface heterogeneity from complicated production signals.
- Physical Constraints in Machine Learning Models – development of methods to encode physical constraints in machine learning methods for improved accuracy and uncertainty models.
- Reduced Dimensionality Representation of Subsurface Uncertainty – use of methods such as autoencoders to extracts salient features from complicated systems. This includes the representation of subsurface uncertainty with reduce dimensionality for improved uncertainty modeling and communication.
- Fast Solutions of Inverse Problems - estimate (with measures of uncertainty) lithology, petrophysical, elastic, and mechanical properties from noisy surface and borehole geophysical measurements. Approximate numerical solution for ultra-fast deep-learning training.
- Rock and Fluid Inference from Petrophysical Measures – modeling of the complex relationships, and data issues to improve inference of rock and fluids from petrophysical measures.
- Seismic Downscaling – modeling the relationships between low resolution seismic and high resolution reservoir features and other multi-scale procedures to relate input measurements to output properties.
- Anomaly Detection – methods to detect features based on concepts such as significance for production and artifacts from spatiotemporal datasets.
- Impact of Well Trajectory on Production - quantify effect / influence of well placement / quality on production to improve well trajectory design.
- Advanced Optimal Well Placement - Real-time integration of sub-surface information (geophysics, geostatistics) and drilling data for placing wells optimally for primary & secondary HC recovery.
- Advanced Optimal Drilling Management - Real-time characterization of pore-pressure, in-situ stress, rock properties, for optimum mud pressure and properties, etc.
- Feature Engineering – methods for encoding spatial and physical constraints while potentially reducing problem dimensionality.
- Building Reservoir Models – machine learning-based reservoir models that honor complicated heterogeneity concepts, seismic information and production data. Integrated approach for estimation of inter-well and missing properties (data imputation).
- Multiscale Flow Proxy Models – fast assessments of flow behavior for model upscaling (from pores scale to production scale). We have already demonstrated the use of convolutional neural nets.
Recent numerical developments and improved computational resources have led to a rapid expansion of big data analytics and machine learning implementations. Oil and gas has a long history with big data from seismic surveys, production monitoring along with various other remote sensing and well-based data, and has developed various physics-based engineering and stochastic statistical workflows. There is an opportunity to combine best-practice and cutting-edge technology in reservoir spatiotemporal characterization and modeling, production data integration, reservoir geophysics and real-time drilling control with big data analytics and machine learning to optimize well trajectory and resource recovery.
Optimum well trajectory and resource recovery through integration of engineering, data analytics, and machine learning. The hydrocarbon industry requires high-resolution, integrated physics-and geology-based and data-driven, real-time updateable models that are cost-effective, interpretable, efficient, and reliable for production-oriented optimum decision-making, for both conventional and unconventional resources.
This DIRECT industrial affiliates program, based in the Hildebrand Department of Petroleum and Geosystems Engineering, at the University of Texas at Austin, will work to develop these integrated modeling and decision support systems to solve the following outstanding problems:
- Integration: Maximizing the integration of deterministic engineering, geological description, target-oriented drilling, geophysical measurements, borehole formation evaluation, production history and core data to construct high-resolution reservoir models for improved production forecast accuracy.
- Characterization: Improving the spatial resolution of reservoir description and modeling based on enhanced data integration for improved development decision-making.
- Grey Box Modeling: Development of big data analytics and machine learning methods that fully account for geospatial and engineering knowledge.
- Robust Decision Making: Automated, expert systems to support consistent evaluation of subsurface and production data.
- System Interpretability: Advanced system summarization and spatial visualization for model interrogation and learning from models for credible decision support.
- Optimum Drilling: Development of modern, production-oriented drilling strategies by designing trajectories for optimum well placement to maximize reserves intersection and recovery factors by primary or secondary production means.
- In-fill Drilling: Development of modern, efficient, and cost-effective strategies to evaluate in-fill drilling, primary or secondary production, and intelligent feedback control systems for reactive production under variable geological, fluid and financial constraints.
- Uncertainty Quantification: Development of modern methods to ascertain the value of measurements and the uncertainty of descriptions and quantifications.
- Modern Software Solutions for Reservoir Characterization: Development of modern computer and software solutions for rapid and efficient 3D collocated multi-physics description, visualization, modeling, well geosteering, and production forecasting.
The faculty leading this IAP are uniquely capable to address these challenges, given our strong knowledge concerning geology, geophysics, geomechanics, drilling and completions, reservoir engineering, formation evaluation, geostatistics, reservoir modeling, data analytics and machine learning.
Consortium start-up requires three supporting partner companies at $60k/year. At this level of support, the consortium will be able to support 3-4 PhD students to conduct the planned research supervised by leading faculty and while integrating input from the consortium participants. Interested companies are welcome to join now. We will host a formal kick-off session in The Woodlands, TX, June 13th 2019.
We are happy to discuss,
Prof. Michael J. Pyrcz, Ph.D., P.Eng. e-mail profile website Twitter GitHub YouTube
Prof. John Foster, Ph.D. e-mail profile GitHub YouTube
Prof. Carlos Torres-Verdin, Ph.D. e-mail profile
Prof. Eric van Oort, Ph.D. e-mail profile
Hildebrand Department of Petroleum & Geosystems Engineering
The University of Texas at Austin
200 E. Dean Keeton St., Stop C0300
Austin, TX 78712-1585