In this paper, we present a water-cut estimator utilizing the function approximation capability of an artificial neural network (ANN). The inputs to the ANN are optical sensor readings in a Red-Eye water-cut meter, which features the near-infrared (NIR) absorption spectroscopy technology. The initial training of the ANNwas done with a data set acquired from our multiphase flow-loop test facility, which was filled with live oil, water and gas. The test fluid stream was adjusted with good ranges of water-cut and gas-volume fractions which were supposed to cover the situations that can be foreseen in real production. However, clear discrepancies between the outputs of the ANN and the water-cut values from BS&W measurmentswere observedwhen the ANN was applied to actual production data measured by Red-Eye meters installed at two offshore wells. To address this issue and equip the ANN with self-adapting capability in real application, we propose a Bayesian approach to update the parameters of the ANN based on both initial flow-loop data and collected field data. The performance of the adapted ANN on both the data sets shows the effectiveness of the method.
In this study, the authors investigated a fully data-driven approach using artificial neural networks (ANNs) for real-time virtual flowmetering and back-allocation in production wells. Australian technology developer MezurX is touting its newly introduced flow, density, and mud monitoring system as a significantly better alternative to the widely used Coriolis meter. Technology developers are working on a new ultrasonic flowmeter for drilling fluids aimed at obtaining better measurements than current systems with far fewer disruptions to rig systems.
In this study, the authors investigated a fully data-driven approach using artificial neural networks (ANNs) for real-time virtual flowmetering and back-allocation in production wells. The authors present a new data-driven approach to estimate the injection rate in all noninstrumented wells in a large waterflooding operation accurately.
High-fidelity 3D engineering simulations are valuable in making decisions, but they can be cost-prohibitive and require significant amounts of time to execute. The integration of deep-learning neural networks with computational fluid dynamics may help accelerate the simulation process. Reducing a separation system’s footprint while increasing separation efficiency is demonstrated in an Oklahoma field trial. Reliable separation is becoming an enabling technology to help develop remote location resources and more difficult applications, such as heavy oil, produced water, sand disposal, and back-produced fluids in enhanced oil recovery. This paper provides details of comprehensive computational-fluid-dynamics (CFD) -based studies performed to overcome the separation inefficiencies experienced in a large-scale three-phase separator.
There is a growing trend toward the use of statistical modeling and data analytics for oil and gas (and related subsurface domain) applications. This session will focus on the state of the art applications of converting data into information—particularly the actionable kind that lead to better decisions.
Operators are looking for ways to better handle water coming from subsea wells, which is typically treated at topside facilities. Subsea separation systems are not equipped to discharge water back into the reservoir, so how do companies close the gaps? High-fidelity 3D engineering simulations are valuable in making decisions, but they can be cost-prohibitive and require significant amounts of time to execute. The integration of deep-learning neural networks with computational fluid dynamics may help accelerate the simulation process. Saudi Aramco studied such algorithms to produce images simulating the flow inside a pipe’s cross section, possibly reducing the need for separator-based multiphase flowmeters.
The course introduction will attempt to answer the question: How will A.I. change the way we work in the Oil and Gas industry in the coming years? Looking at what is underway in other industries and guessing what type of projects are under development in R&D departments in our industry will help answer that question. Oil and Gas examples will be presented corresponding to each of the terms A.I., Machine Learning, and Deep Learning, allowing participants to reach a clear understanding on how they differ. The course will then focus on Deep Learning (DL) and address all key aspects of developing and applying the technology to Oil and Gas projects. What is DL and how different is it from traditional neural networks?
At times, it may seem that machine learning can be performed without a sound statistical background, but this does not take in to account many difficult nuances. Code written to make machine learning easier does not negate the need for an in-depth understanding of the problem. Nicknamed “warshipping,” the hacking technique allows remote infiltration of corporate networks by hiding a remote-controlled scanning device designed to penetrate a wireless network inside a package.
Expert-guided machine learning has been used to classify depositional facies from core photographs of the Wolfcamp, Bone Spring and Spraberry formations in the Permian Basin. Training sets of core facies were selected by a sedimentologist. A model was built using a convolutional neural network and then tested against core outside of the training set with a 98% accuracy. The system can yield a quit-look of core facies much faster than that of traditional methods.
Artificial Intelligence (AI) is a branch of computer science that creates intelligent machines that work and react like humans. Machine learning is a key part of AI and requires an ability to identify patterns in streams of inputs. Learning with adequate supervision involves classification, which determines the category an object belongs to. Today it is being extensively used in image and speech recognition. At present the application of machine learning is in its infancy in the area of geosciences for the oil and gas industry.
The objective of our research is to determine if machine learning can be used to fast-track identification of depositional facies from images of conventional core photographs. Normally, this work requires a sedimentologist to painstakingly describe a core that may take many weeks to incorporate with logs and other formation evaluation data. With over 200 cored wells having 50,000 feet of core in our Permian Basin projects, the task of core description is overwhelming.
Theory and/or Methods
In order to meet the objective the software needs to be trained to recognize the various depositional facies. This is done by employing a sedimentologist (expert) to guide the training with the AI specialist. The sedimentologist builds a training set from several cores through a particular formation (e.g. Wolfcamp). The training set is a set of images selected by the sedimentologist to cover the range of depositional facies and the variations seen in each facies (Figure 1). The training sets typically employ 20 to 40 images of each facies.
Pore pressure prediction plays a critical role in the ability to predict areas of high overpressure and fracture behavior for the exploitation of unconventional plays, which are both correlated with production. Shales in these plays have variable clay content and complex multi-mineral fractions that require a detailed petrophysical assessment reinforced with rock physics modelling as needed. For example, changes in total organic content have a similar elastic response to changes in porosity. Therefore, any pressure-stress property model for unconventional plays must be supported by petrophysically conditioned elastic logs and accurate multi-mineral volume sets calibrated to core data.
A supervised deep neural network approach is introduced as an alternative innovative tool for petrophysical, pore pressure and geomechanics analysis enabling the use of all the previously collected and interpreted data to devise solutions which simultaneously integrate wide ranging well bore and wireline logs. We implement three neural networks, all with similar structure, as each of these networks had a different objective and the outputs from one were the inputs for the other.
The first network was trained to predict petrophysical volume logs (shale, sand, dolomite, calcite, kerogen and also porosity) simultaneously from compressional velocity (Vp), Gamma ray, density (rho), resistivity and Neutron logs. The second neural network, cascaded from the first, was then designed to match the manually predicted pore pressure. The inputs were Vp and shear velocity (Vs), Rho, resistivity, Neutron logs as well as the results of the first network. The third network focused on predicting various properties of interest, in this case pore pressure, minimum horizontal stress (Shmin), maximum horizontal stress (SHmax), and volume of kerogen, based on only Vp, Vs, and Rho logs which is an example building a neural network capable of predicting key rock properties directly from seismic inversion results to produce meaningful 3D interpretations.
The volumetric pore pressure model was also positively correlated to cumulative production values from blind long horizontal wells. The results show a promising outlook for the application of deep learning in integrated studies such as those shown in this paper.