2019 Spring Workshop

A Machine Learning Based Automated System for Well Log Editing

Ridvan Akkurt (Schlumberger)

Recent advances in data science and machine learning (ML) have brought the benefits of these technologies closer to the main stream of Petrophysics. ML systems, where decisions and self-checks are made by carefully designed algorithms, in addition to executing typical tasks such as classification and regression, offer efficient and consistent solutions to the modern petrophysicist. The outline of such a system and its application in the form of a multi-level workflow to a 59-well multi-field study are presented in this paper.

The main objective of the workflow is to identify outliers in bulk-density and compressional slowness logs, and to edit (reconstruct) them using data-driven predictive models. The system is fully automated, designed to optimize the use of all available data, and provide uncertainty estimates. It integrates modern concepts for novelty detection, predictive classification and regression, as well as multidimensional scaling based on inter-well similarity.

The original motivation for the workflow lies in data-room applications where one has to quality control and evaluate large amounts of data in a few days, often with limited opportunities for validation or cross-checking.  A workflow designed for such applications must not be only automated, but also flexible in terms of its input requirements and guidance from the user. Our  workflow does not require any supervision,  well locations or formation tops are not needed, and it can work without the usual quality control inputs such as caliper or density-correction logs.

Benchmarking of ML results against those created by human petrophysical experts show the ML workflow can provide high quality answers that compare favorably to those produced by petrophysical experts. A second validation exercise, that compares acoustic impedance logs computed from ML answers to actual seismic data, provides further evidence for the efficiency of the ML generated results.

The ML system supports the petrophysicist by easing the burden on repetitive and burdensome quality control tasks. The efficiency gains and time savings created can be used for enhanced effective cross-discipline integration, collaboration and further innovation.

Successful Multivariate Modeling of Production from Oil and Gas Wells Enabling the Testing of Significant Managerial Questions

Richard Batsell (Rice University)

This presentation extracts material from a joint paper with Sanjay Paranji (Anadarko) and Jason Mintz (Apache) first presented in a July, 2018 URTeC Conference.  The presentation begins by using data from 452 gas wells drilled over a recent 4 ½ year period in a Northeastern State to demonstrate the development of successful multivariate models explaining 60% to 70% of the variance in production.  Reflecting both geology and completion attributes, the models reveal the relative importance of various key variables as well as how the importance changes as the production accumulates.  And, because they control for both geology and completion variables, the models also allow for statistical tests comparing:  1) Production from different formations where costs to reach the formations are quite different (thus saving millions of dollars); 2) Linear versus non-linear effects of lateral length; 3) Competing fracking approaches offered by major suppliers; and, several other such tests.  The general approach, using this particular data set as an example, can be used to objectively and concretely address a large number of critical managerial decisions.  Although the data used in this presentation come, specifically, from gas wells, similar successful applications have been achieved in fields in the Marcellus (Pennsylvania), the Eagle Ford (Texas), and the Wattenberg (Colorado) to name a few.

Petrophysical Multimineral Analysis Using a Genetic Algorithm: Applications to unconventional reservoirs

Reinaldo Michelena (SeisPetro)

Petrophysical multimineral analysis is a tool that can help relate the complexity of the rock composition in conventional or unconventional reservoirs to the well log measurements by assuming the log response is a linearized combination of individual constituents of known properties or “end-points”. Even though the assumption of linearity of the log response is usually valid, the assumption of known properties is typically the exception rather than the norm in unconventional reservoirs due to the complexity of the mixtures, inconsistencies in tabulated properties of the same mineral, uncertain kerogen properties, and scarce or unreliable measurements of water salinity. Often, due to the limited number of logs available, the petrophysicist is forced to create artificial, composite minerals (i.e., “matrix”) whose effective properties are unknown. The fact that the fractions of constituents and their properties are both unknown makes of the current practice of multimineral analysis an “educated” guess exercise with no guarantee of obtaining usable, consistent results.

In this work, we pose the estimation of the constituents and their properties using multimineral analysis as a stochastic nonlinear optimization problem where a genetic algorithm substitutes the time-consuming, manual trial-and-error process of adjusting properties and fitting the input logs. The method still requires interpretative inputs based on prior knowledge and experience, but such inputs are provided in the form of ranges instead of single property values, facilitating the work of the analysist. By testing adaptively thousands of solutions and considerably reducing the time needed to fit the input logs with a consistent set of properties, it becomes then possible to test other scenarios of input data and constituents, quantify the uncertainty and non-uniqueness of individual parameters, and shed light upon higher level petrophysical questions such as spatial variations in water resistivity, kerogen maturity, or clay composition.

We show examples of application of this methodology to three unconventional reservoirs that illustrate how to solve for variable lithology and effective hydrocarbon pore volumes.

Machine Learning for Improved Directional Drilling

Hani Elshahawi (Shell)

Directional drilling is a complex process involving the remote control of tool alignment and force application to a very long drill string subject to variable external forces.  Controlling bit tool face orientation while enduring adequate rate of penetration (ROP) is quite challenging, with aspects that have been described as more art than science.  Improving this control helps preserve proper well trajectory and eliminate deviations that require corrective measures and add to well costs.

An artificial intelligence system was developed to learn from the actions of expert direction al drillers and the mechanics of drilling simulations.  Machine learning algorithms were employed to improve the efficiency of directional drilling: optimized ROP, less tortuous borehole, less personnel on board (POB) and consistency across operations.  The system ingests historical and simulation data corresponding to the information used and actions taken by expert directional drillers and uses that data to generate decisions that result in efficient slide drilling.

To create a system for controlling tool face angle and guiding drill bit sliding during directional drilling, relevant historical data from directional drilling operations was gathered.  Much of this data was recorded in the drilling logs, which the drilling operator traditionally uses to control drilling parameters.  The collected data was then filtered and used to structure and train artificial neural networks and select appropriate hyperparameters.  Reinforcement learning methods were used to refine the neural networks trained on historical data.  A computational model for drill string physics was used to simulate the mechanics of directional drilling.  A successfully trained network was considered one that minimized deviation from planned wellbore trajectory, minimized tortuosity, and maximized ROP.

The neural network developed could replicate the decisions of expert directional drillers within a small error (<3%).  Reinforcement learning was then successfully used to improve network performance, particularly for conditions not previously considered.

Since the algorithm has demonstrated competence in the historical and simulated realms, it will be further tested as a real-time advisory system for control of directional drilling operation.  Ultimately, the algorithm can be directly integrated into drilling operations, enabling fully automated directional drilling.

Machine Learning Applications in Reservoir Characterization

Siddharth Misra (University of OK)

Dr. Misra will present few case studies on the use of machine learning techniques to assist pore scale characterization, geomechanical characterization, and image segmentation. Nuclear Magnetic Resonance (NMR) log acquired in geological formations contains information related to fluid-filled pore volume, fluid phase distribution, and fluid mobility. Deployment of NMR logging tool is financially and operationally challenging in comparison to conventional logging tools, especially in shale wells. To that end, we develop neural network models to generate NMR T2 distribution in the absence of NMR logging tool. In the first case study, we developed one shallow-learning and one deep-learning neural network models that process 10 conventional, easy-to-acquire logs to generate the NMR T2 distributions along 300-feet depth interval of a shale reservoir. In the second case study, we propose a data-driven technique to generate compressional and shear travel time logs (DTC and DTS, respectively) in the absence of sonic logging tool. DTC and DTS logs acquired using sonic logging tools are used to estimate connected porosity, bulk modulus, shear modulus, Young’s modulus, Poisson’s ratio, brittleness coefficient, and Biot’s constant for purposes of geomechanical characterization. Six shallow learning models process 13 conventional and easy-to-acquire logs, namely Lithology, GR, DCAL, DPHZ, NPOR, PEFZ, RHOZ and RLA0-5 logs. Artificial Neural Network model performs the best among the six models. ANN-predicted DTC and DTS logs have R2 of 0.87 and 0.85, respectively, for Well 1. In the third case study, machine learning assisted the segmentation of SEM images of shale samples. The newly developed machine-learning-assisted image segmentation method was applied on several SEM images of shales, and it successfully identified four rock components, namely (1) pore/crack, (2) pyrite, (3) matrix comprising clay, quartz, and cements, and (4) kerogen/organic components. This segmentation method involves two steps, feature extraction from SEM images followed by random forest classification of each pixel in the SEM image.

Geology at the Crossroads of the Future

Matt Belobraydic (Schlumberger)

"With their four-dimensional minds, and in their interdisciplinary ultra-verbal way, geologists can wiggle out of almost anything." – John McPhee

As the oil and gas industry moves to more data driven solutions through big data, cloud solutions, and artificial intelligence, geoscientists are poised to step deeper into the lead integrator role.  Combining different scales, vintages, and sources of data is a requirement to maximize ROI in oil and gas fields and plays.  Gone are the days of siloed teams.  With cheap data storage and faster model realizations, multiple working hypotheses can be tested utilizing multidomain interpretations that can be integrated back into analyses, creating a positive feedback loop to identify true play and basin drivers, quantify uncertainties, and minimize risk.

Cloud computing, artificial intelligence, and new correlation methods are making it easier to find well locations, targets, and design completion strategies that provide the most economic advantageous way to extract hydrocarbons.  Data scientists are creating new ways to make tedious parts of interpretations more automated, leading to a larger amount of data available for incorporation into final analyses.  Correlations and results, however, may not make sense without being "ground-truthed" with real world geologic knowledge.

Through integrated teams and the increase in available data and interpretations, geologists are in a unique position to "wiggle" into the role of leading the data science revolution currently underway in the petroleum industry.  Using the Bakken and Three Forks plays in the Williston Basin as an example; the geologic domain as an integration platform for petrophysics, geomechanics, production and stimulation engineering, reservoir engineering, management, and (of course) geology will be demonstrated.

Geosteering an Unconventional Shale Lithozone with Confidence

Kim McLean (Emerson/Paradigm)

Unconventional shale reservoirs will continue to be a focus of the energy industry in North America for the foreseeable future. The bulk of the costs incurred in producing shale reservoirs comes from frac design and completions. With the price of oil still somewhat volatile, it is in the interest of energy companies to use data on hand to plan their horizontal wells such that they can stay in zone, and ultimately complete the wells in the optimal zone of interest.

Shale reservoirs tend to be heterogeneous in nature, with facies that differ in mineralogy, and as a result, geomechanical properties. What if a company could integrate knowledge of the geomechancial rock properties from their pilot wells into their well planning?

We evaluate the mineralogy, lithology and mechanical rock properties of an offset well, and use Multi-Resolution Graph Clustering (MRGC) to model mineralogical and mechanical facies in both the offset well and the proposed well paths. We then go a step further to integrate our mineralogical and geomechanical facies using MRGC to create a litho-mechanical facies and evaluate the facies to determine which, if any, would be a 'sweet-spot' for best original oil in place.

We present a workflow that evaluates the mineralogy and geomechanical rock properties of an offset well, and incorporate that information into the planning stages of a horizontal well. Ideally, the geologist could take advantage of their knowledge of geomechanical facies and optimally place a well such that it leads to better placement within the reservoir sweet spot.

Integrated stochastic workflow for optimum well spacing with data analytics, pilots, geomechanical-reservoir modeling, and economic analysis

Richard Cao (Shell)

Appropriate well spacing decision is critical for unconventional asset development. Operators generally use three methods to identify optimum well spacing: trial and error, pilots, and modeling. The field trials and pilots often involve significant capital investment and require significant time before conclusive results can be observed. On the other hand, modeling is cost effective and time efficient but has large uncertainties in the results, which must be carefully calibrated with the field measurements to narrow down the uncertainty range. In this paper, these three methods are systematically integrated to study well interference and identify optimum well spacing for Wolfcamp development at Delaware basin.

First, multiple well spacing pilots are drilled and produced with various diagnostic signals collected, such as Microseismic, Bottom Hole Pressure (BHP), fluid sampling, etc. Then the general production trend analysis is performed for these pilot results combined with other available public data from the Delaware basin. After that, a well spacing trial with good quality of data is selected as the modeling target. A full-scale 3D multi-well reservoir model with geomechanical effects was built for history match of the oil, gas, water productions, and pressure. The modeling results are carefully calibrated with the field data at different time and length scales. The potential production interference of different well spacing is captured as Estimated Ultimate Recovery (EUR) reduction compared to the base case. At the same time, stochastic studies with hundreds of simulation runs are performed on a simple reservoir model to investigate the impacts of production of interference for different well spacing and subsurface parameters.

The results show that the production interference among horizontal wells have large uncertainty due to the heterogeneous subsurface parameters and the hydraulic fracturing process. The EUR reduction correlations obtained from this workflow then integrate with economic criteria (i.e. capital efficiency, Net Present Value (NPV), and cash flow) to determine the optimum well spacing.

Rock Typing in Organic Shales: Barnett, Eagle Ford, Woodford and Wolfcamp

Ishank Gupta (University of OK)

In this work, an integrated workflow is presented for rock typing using lab petrophysical measurements, logs, and production data. The key petrophysical parameters used for rock typing are porosity, total organic carbon (TOC), mineralogical compositions and mercury injection capillary pressure (MICP). Principal Component Analysis (PCA) is used to reduce dimensionality of the dataset and improve efficiency of the clustering algorithms. Unsupervised clustering algorithms like K-Means and Self Organizing Maps (SOM) are used to define rock types. The integrated workflow is applied separately for four shale plays namely Barnett, Eagle Ford, Woodford and Wolfcamp.

A total of 25 wells with core data are considered for rock typing in the four shale plays. The rock types are upscaled to more than 140 wells representing a 20,000-ft. depth interval. A manual approach would have been prohibitively time-consuming.

Rock Type 1 is generally characterized by high porosity, high TOC, and high brittleness. Not surprisingly, Rock Type 1 has the highest positive impact on well productivity. Rock Type 2 has intermediate values of porosity and TOC and thus, moderate source potential and storage. Rock Type 3 has the highest carbonate content (except Eagle Ford) and poor storage (except Eagle Ford) and source rock potential.

Classification algorithms like Support Vector Machines (SVM) are used to upscale rock types from core data to logs. The training datasets comprise of depths at which both core and log data are available. Different logs like gamma ray, resistivity, neutron porosity and density are used for upscaling. Finally, a rock type ratio (RTR) is defined based on the fraction of Rock Type 1 over gross thickness. RTR is found to strongly correlate with normalized oil equivalent production rate.

Automatic Production- and Fabric-Oriented Rock Classification and Reservoir Evaluation in Organic-Rich Mudrocks through Integration of Multi-Scale and Multi-Physics Formation Data

Zoya Heidari (University of Texas at Austin)

Reliable reservoir characterization through integration of multi-scale and multi-physics formation data is essential for successful production in complex formations such as organic-rich mudrocks. Complex rock fabric, composition, pore structure, rock mechanics, and geochemistry introduce uncertainties in rock physics models used for formation evaluation, which can significantly affect outcomes of rock classification efforts for detecting the best landing spots and completion intervals. Furthermore, automatic integration of multi-scale and multi-physics formation data such as image logs, conventional well logs, and core measurements for formation evaluation and rock classification can be challenging, especially in the absence of reliable and sufficient data.

In this presentation, the impacts of choice of data and analysis methods on rock classification and its influence on production decisions will be discussed. Automatic, intelligent, integrated, and production-oriented rock classification workflows will be introduced, which simultaneously quantify rock fabric from image logs (and/or core images), optimize the number of rock classes (without a priori knowledge of rock types) and enhance reservoir evaluation and production decisions.  Outcomes of recent research developments in our team including field applications (e.g., application to the Midland Basin) will be presented. Results demonstrate that decisions made for landing spots and completion intervals in organic-rich mudrock formations can be enhanced by reliable rock classification that quantitatively takes into account rock fabric, reservoir petrophysics, rock mechanics, and geochemistry through integration of multi-scale and multi-physics formation data.