Machine Learning – digmart.io

Artificial Intelligence

Artificial intelligence (AI) is a computer system trained to perceive its environment, make decisions, and take action. AI systems rely on learning algorithms, such as machine learning and deep learning, along with large sets of sensor data with well-defined representations of objective truth.

Machine learning: We use machine learning as shorthand for “traditional machine learning”—the workflow in which you manually select features and then train the model. When we refer to machine learning we exclude deep learning. Common machine learning techniques include decision trees, support vector machines, and ensemble methods.

Deep learning: A subset of machine learning modeled loosely on the neural pathways of the human brain. Deep refers to the multiple layers between the input and output layers. In deep learning, the algorithm automatically learns what features are useful. Common deep learning techniques include convolutional neural networks (CNNs), recurrent neural networks (such as long short-term memory, or LSTM), and deep Q networks.

High-Def Rock Properties in 3D

Reservoir characterization is an ambitious challenge that aims to predict variations within the subsurface using fit-for-purpose information that follows physical and geological sense. To properly achieve subsurface characterization, artificial intelligence (AI) algorithms may be used. Machine learning (ML), a subset of AI, is a data-driven approach that has exploded in popularity during the past decades in industries such as healthcare, banking and finance, data security, and e-commerce. An advantage of machine learning methods is that they can be implemented without knowledge of theories and equations backed by science. The principal challenge of machine learning lies in attaining enough training information, which is essential in obtaining an adequate model that allows for a prediction with a high level of accuracy.

Our comprehensive machine learning console is used to quickly and easily predict rock properties for reservoir characterization studies to reduce subsurface uncertainty. We accomplish this by integrating seismic and well data to predict high frequency acoustic, elastic and petrophysical properties in three dimensions. Our technology is backed by advanced machine and deep learning algorithms using ensemble learning and neural networks.

The figure below is from a siliciclastic reservoir mapping case study showing a High-Def gamma ray random line result between two wells. Well A was used to build the machine learning model. Well B was only used to build the compaction trend. The base of the reservoir sand is also indicated. A 120 Hz cap has been applied to the measured logs. It must be noted that previous attempts to map this sand using conventional seismic inversion proved to be unsuccessful. The data are provided by the U.S. Department of Energy’s (DOE) National Energy Technology Laboratory (NETL).

High-Def 3D gamma ray random line. A 120 Hz cap has been applied to the measured gamma ray logs.

The next figure compares the High-Def gamma ray with fullstack seismic, which does not have enough frequency content to adequately map the sand formation. Notice the increased level of layer detail in the gamma ray result throughout the analysis window, which is validated by the match with the measured logs.

Comparison of conventional seismic (top) and High-Def gamma ray. The reservoir sand is indicated by a white arrow. Well A was used to build the machine learning model. Well B is a blind well. The measured logs have a 120 Hz filter applied.

Log Prediction

In seismic reservoir characterization, sonic measurements must be available at the wells that will be used to map the lateral extent of the reservoir. The acquisition and processing of sonic data is costly and can easily increase the required budget in a reservoir characterization campaign. Machine learning can help significantly reduce the need to pay for sonic logs.

The figures below are from a sonic and shear sonic prediction case study using existing gamma ray and density logs to train the machine learning model. The learning well is used to build the model. The application well prediction is completely independent of the well information used to build the model. Notice the 85% correlation for DTCO prediction and 83% correlation for DTSM. Histogram and crossplot comparisons between the measured and predicted logs are shown for quality control measures.

Machine learning (ML) compressional sonic log prediction using GR and RHOB as input. The learning well is used to build the ML model. The application well is used solely for quality control and has no input into the ML model. The quality of the result is evaluated using the R2 method and has a correlation of 85%.

Machine learning (ML) shear sonic log prediction using GR and RHOB as input. The learning well is used to build the ML model. The application well is used solely for quality control and has no input into the ML model. The quality of the result is evaluated using the R2 method and has a correlation of 83%.

Classification

Bayesian decision theory allows for expressing the probability of a particular class given an observed x as

P(Cj│x) = P(x,Cj)/P(x) = (P(x│Cj) P(Cj))/P(x)

where P (x, C_j) is the joint probability of x and C_j, P (x | C_j) denotes the conditional probability of x given C_j. In order to establish a lithoclass x (i.e. gas sand, oil sand, brine sand, shale, etc.), a rock diagnostic is performed where, using the available well log data, a set of crossplots is generated utilizing the inversion output attributes (i.e. AI, SI, V_p/V_s, Poisson’s ratio, Lambda-rho, Mu-rho, etc.) and petrophysical logs (i.e. VCL, PHIT, S_w, etc.). The objective is to determine which elastic attributes best separate out the desired lithoclasses in the crossplot analysis. Last, probability density functions (pdfs) are generated from the well log data and are applied to the selected seismic inversion output cubes to produce class and probability cubes. At each sample, there is an associated probability which represents the likelihood of each class defined based on the cross plot analysis explained above. Several examples as well as a more comprehensive explanation of Bayesian statistics is given in the book Quantitative Seismic Interpretation by Avseth et al. (2005).

The next figure shows a map view of pay sand probability horizon slice. The result is part of a reservoir characterization study in the Burgos Basin, which includes a spatial lithology distribution based on lithoclasses generated using rock physics and Bayesian statistics. The lithology prediction was preceded by prestack simultaneous AVO inversion. The study was carried out to determine the lateral extent of the producing gas-saturated sand which has been penetrated twice successfully (Wells A and B) based solely on seismic amplitudes. Well C was a dry hole, which inspired the operator to find a different methodology to accurately characterize the reservoir sand.

The result may be interpreted as: given the information known from wells A and B, there is an 88.0% probability that hydrocarbons exist at Well A, which is where we know the well produced gas; the same goes for Well B, which shows a 61.5% probability. In Well C, we know there are no hydrocarbons and has a 47.6% probability. One may assign a conservative threshold of ~55% – 60% to determine whether gas saturated sand is present or not.

Map view of pay sand probability horizon slice.

The next case study is an example of using Bayesian statistics to reduce the risk of drilling. The patented workflow is discussed in The Interpretation Journal – November 2013, by Leiceaga et al. In this study, seismic-derived elastic properties from simultaneous inversion were successfully correlated with hydrocarbon production in an unconventional reservoir using Bayes’ theorem. A production classification was carried out where all wells were split into three categories (high, medium, and low) with respect to gas production. The classification was based on a 30-day average, initial production (IP) rate, where a medium producer fell within 3–5 MMCF∕day. Moreover, production rates are normalized to the lateral length of the horizontal wells.

The color scheme used in the resulting hydrocarbon production capacity (HPC) volume follows a traffic-light approach where green corresponds to expected high production, yellow to medium, and red to low; nonclassified samples are indicated in black.

The figure shows a hydrocarbon production capacity section at a vertical well where the upper perforation gave high production (green) while the lower perforation set is low (red). The result shows a good correlation with the observed production in both perforated intervals. The result also shows potential zones of high production capacity downdip.