10 Applications of Machine Learning Within Petrophysics

Examples of machine learning to enhance your petrophysical workflow

Several decades of hydrocarbon exploration have led to the acquisition and storage of large quantities of well related measurements, which have been used to characterise the subsurface geology and its hydrocarbon potential. The potential of these large volumes of data has been increasingly exploited over the last couple of decades as computational power and the adoption of new machine learning algorithms increase. Within the petrophysical domain, machine learning has been used to speed up workflows, characterise the geology into discrete electrofacies, make predictions, and much more.

The content within this article was presented in September 2021 at the London Petrophysical Society, a modified version of that presentation is available on my YouTube channel. Make sure to like and subscribe to it for more content like this.

What is Petrophysics?

Petrophysics is a discipline that studies the physical and chemical properties of rocks and their interactions with contained fluids. It is a major discipline within the oil and gas industry and aims to understand and quantify the properties and quantities of moveable hydrocarbons within a reservoir. Derived properties include porosity (the amount of storage space available to store hydrocarbons), permeability (how well the pores are connected with each other), and water saturation (the amount of pore space filled with water, which can, in turn, be used to calculate hydrocarbon saturation.

Petrophysical projects involve analysing measurements that have been obtained by downhole logging tools. These tools measure various properties of the rocks including their natural radioactivity, and their response to electrical currents, live radioactive sources, and acoustic signals. Many of the properties that are derived are subsequently transformed and used within well-established empirical equations (e.g. the Archie Equation) to calculate the required petrophysical properties of the reservoir. Many authors have attempted to bypass the empirical equations and build machine learning models that utilise the well log data to directly predict the petrophysical properties.

Additionally, as with any project that involves data, a significant amount of time (up to 90%) can be spent on tedious, but very important tasks such as data gathering, data quality control, data corrections, and data transformations. As such, many articles and applications of machine learning focus on improving the efficiency and speed of these stages.

The following article touches on 10 examples where machine learning has been used to help with various aspects of the petrophysical workflow. Each example contains a list of references to key and interesting papers where these techniques have been employed.

1. Automated Outlier Detection

Outliers are data points that sit outside of the normal or expected statistical distribution of the dataset and they can occur for a variety of reasons within well log measurements. Reasons include, but are not limited to:

Measurement and sensor errors
Borehole washout
Drilling vibrations impacting Logging While Drilling tools
Unexpected events and geology

It is essential that outliers are identified and handled appropriately. This can be achieved through manual methods such as using Z-Scores, boxplots, and conventional crossplots (scatterplots).

A number of research papers recently have focused on the use of unsupervised outlier detection algorithms to highlight potential outliers. These methods include K-Means Clustering, Angular Based Outlier Detection, Support Vector Machines, and Isolation Forests.

To fully confirm if a point or set of points are outliers, domain knowledge should be used. This will reduce the risk of misclassification of points that may be real data.

Articles for further reading:

Akkurt et al., (2018): Accelerating and Enhancing Petrophysical Analysis With Machine Learning: A Case Study of an Automated System for Well Log Outlier Detection and Reconstruction
Banas et al., (2021): Novel Methodology for Automation of Bad Well Log Data Identification and Repair
Misra et al., (2020): Unsupervised outlier detection techniques for well logs and geophysical data

2. Well Log Repair

Once outliers and bad data have been identified they can usually be removed or repaired prior to carrying a petrophysical interpretation or applying a machine learning model.

The process of identifying and checking data quality issues can take up a significant amount of time in a project, as such numerous authors have looked at ways to automate / semi-automate this process. Models are trained on sections of the well that are considered “good”, and then used to predict over the interval with issues.

Articles for further reading:

Banas et al., (2021): Novel Methodology for Automation of Bad Well Log Data Identification and Repair
Cuddy (2020): The Benefits and Dangers of Using Artificial Intelligence in Petrophysics
Cuddy (2002): Using Fuzziness to Repair Borehole Electrical Logs
Singh et al., 2020: Machine Learning Assisted Petrophysical Logs Quality Control, Editing and Reconstruction

3. Well Log Normalisation

Well log normalisation is a common part of the petrophysical workflow. It is used to remove data inaccuracy caused by a number of issues including, different tool and sensor technologies, differences in the borehole environment, and issues with tool calibrations.

Normalisation is the process of re-scaling or re-calibrating the well logs so that they are consistent with other logs in other wells within the field or region. This can be manually achieved by applying a single-point normalization (linear shift) or a two-point normalization (‘stretch and squeeze’) to the required curve.

Normalization is commonly applied to gamma-ray logs but can be applied to neutron porosity, bulk density, sonic and spontaneous potential logs as well. Resistivity logs are generally not normalized unless there is a sufficient reason to do so (Shier, 2004).

Articles for further reading:

Akkurt et al., (2019): Machine Learning for Well Log Normalization
McDonald (2021): Petrophysics: Gamma Ray Normalization in Python
Shier (2004): Well log normalization: Methods and guidelines

4. Automated Well Log Correlation

Well to well correlation is a common process where one or more geological intervals are linked together based on geological similarities, such as lithology, biological contents, or mineralogy. This can be a time-consuming and laborious task for the geologist and many authors have attempted to automate this process since the 1970s (Brazell et al., 2019). More recently, machine learning has been used to carry out well log correlations with a high degree of success.

Even though well to well correlation is usually carried out by geologists, there are occasions when a petrophysicist has to work with tops and correlations. When carrying out petrophysical analysis, similar parameters are often used across different wells for the same formation. Ensuring that well logs are correctly correlated allows this process to be carried out more smoothly.

Articles for further reading:

Bakdi et al., (2020): Automated Well Correlation using Machine Learning and Facial Recognition Techniques
Brazell et al., (2019): A Machine-Learning-Based Approach to Assistive Well-Log Correlation
Tokpanov et al., (2020) Deep-learning-based Automated Stratigraphic Correlation

5. Prediction of Missing Well Logs

It is essential that datasets are as complete as possible when carrying out subsurface characterisation, however, there are situations where well logging measurements are missing. This can include different vintages of well logs, logging speeds exceeding tool sampling rates, borehole environmental issues, improper data management, expensive acquisition costs, and casing effects.

In these situations, data can be infilled using empirical relationships between the remaining logging measurements or by using machine learning models from the current well and/or offset wells.

Articles for further reading:

Akinnikawe et al., (2018): Synthetic Well Log Generation Using Machine Learning Techniques
Anemangley et al., (2018): Machine Learning Technique for the Prediction of Shear Wave Velocity Using Petrophysical Logs
McDonald (2020): Visualising Well Data Coverage Using Matplotlib

6. Depth Alignment of Logging Measurements

Depth is a critical measurement for the successful development and completion of a well, with all well logs being tied to it. It allows the selection of perforation intervals, and the setting of packers at the correct depth. Therefore, it is essential that all measurements are using a consistent depth reference.

Well log measurements can be off depth with each other for a number of reasons including different depth references between wireline and logging while drilling passes, cable stretching, different sampling rates between tool types and logging passes, and even variations in weather conditions and sea surface swell on semi-submersible drilling platforms. Ensuring that well logs from multiple passes and runs has been a longstanding challenge within the oil and gas industry.

Alignment of data has traditionally been carried out manually, whereby two or more logging curves are compared from multiple passes and pins added at key features. These features are then aligned to create data that is on depth with each other. Often this process can be subject to bias and can be time-consuming to carry out.

Semi-automated and automated approaches have been developed over the years that utilise cross-correlation and covariance measures between two logging curves. However, recently a number of authors have attempted to use machine learning models to automate this process and remove any bias.

Articles for further reading:

Bittar et al, (2020): Reinforced learning technique for multi-well logs depth matching yield better reservoir delineation
Le et al, (2019): A Machine-Learning Framework for Automating Well-Log Depth Matching
Zimmerman et al., (2018): Machine-Learning-Based Automatic Well-Log Depth Matching

7. Prediction of Continuous Curves from Discrete Core Measurements

One of the key deliverables from a petrophysicist is permeability. It is used to provide an indication of how easy fluids (in particular hydrocarbons) can flow through a rock or a reservoir. At present, there is no logging tool that can directly measure formation permeability downhole, instead, it has to be estimated from tool responses or empirical relationships. Many of the empirical relationships have been derived from core measurements in specific geographical and geological areas and may not be fully applicable to other areas. As such, it is common practice to derive a relationship between core porosity and core permeability and apply that relationship to log-derived porosity.

Core data is not always available and may only be present within a handful of wells within a field or development. This is due to coring operations being expensive and time-consuming. As such, multiple authors have employed machine learning models that are trained on key wells and are then used to predict permeability in all other wells.

Articles for further reading:

Akinnikawe et al., (2018): Synthetic Well Log Generation Using Machine Learning Techniques
Al-Anazi and Gates (2010): Support Vector Regression for Permeability Prediction in a Heterogeneous Reservoir: A Comparative Study
Arkalgud et al, (2019): Domain Transfer Analysis — A Robust New Method for Petrophysical Analysis
Elkatatny et al., (2018): New Insights into the Prediction of Heterogeneous Carbonate Reservoir Permeability from Well Logs Using Artificial Intelligence Network
McDonald, 2020: Porosity-Permeability Relationships Using Linear Regression in Python
Saputelli et al., (2019): Deriving Permeability and Reservoir Rock Typing Supported with Self-Organizing Maps (SOM) and Artificial Neural Networks (ANN) — Optimal Workflow for Enabling Core-Log Integration

8. Prediction of Facies

Understanding the subsurface lithology is an important task in geoscience and petrophysics. Deriving a lithology flag or mineral volumes is one of the tasks often assigned to the petrophysicist.

Machine Learning algorithms have routinely been adopted to group well log measurements into distinct lithological groupings, known as facies. This process can be achieved using either unsupervised learning or supervised learning algorithms.

Unsupervised clustering of data is a common form of exploratory data analysis (EDA) which is used to divide up the data into different groups based on shared characteristics or properties. Data points that are similar to each other are grouped together in the same cluster, and those that are different are placed in another cluster.

Articles for further reading:

Cuddy & Putnam, 1998: Litho-facies and permeability prediction from electrical logs using fuzzy logic.
Hall, 2016: Facies Classification Using Machine Learning
McDonald, 2021: How to Use Unsupervised Clustering on Well Log Data With Python
Zhang & Zhan, 2017: Machine Learning in Rock Facies Classification: An Application of XGBOOST

9. Geomechanical Property Prediction

Well log measurements, such as bulk density and acoustic slowness can be used to gain insights into the stress and rock properties of a reservoir (Young’s Modulus, Bulk Modulus, Poisson’s Ratio, etc). The derivation of geomechanical properties is an essential step in well in planning to ensure wells are drilled safely and completed successfully. It also allows an understanding to be obtained of how the rocks will change and deform in response to changes in stress, temperature, and pressure.

The most accurate way to obtain geomechanical properties is from destructive testing of core samples obtained from the well. Often this is done on a small number of samples and is used to calibrate and verify empirical relationships between log measurements and the properties. When the core samples are obtained in conventional reservoirs, it is usually over the reservoir section, with the overlying intervals ignored. This leads to extrapolating the relationships to these sections if they are of interest.

Articles for further reading:

AlBahrani et al., (2021): Building an Integrated Drilling Geomechanics Model Using a Machine-Learning-Assisted Poro-Elasto-Plastic Finite Element Method
Song et al., (2021): Prediction and Analysis of Geomechanical Properties of Jimusaer Shale Using a Machine Learning Approach

10. Petrophysical Property Prediction

Determination of key reservoir properties is one of the main tasks for a petrophysicist and is normally achieved through well-developed relationships and equations. However, a number of studies have focused on the prediction of these properties using well log measurements.

Articles for further reading:

Khan et al., (2018): Machine Learning Derived Correlation to Determine Water Saturation in Complex Lithologies
Basu et al., (2020): Petrophysical Workflows vs Machine Learning — A Comparative Case Study in the Dakota Group, Williston Basin
MacGregor et al., (2018): Streamlining Petrophysical Workflows with Machine Learning

Summary

There we have it! We have seen 10 examples of where machine learning can benefit and add value to a petrophysical workflow. These 10 examples are by no means the only ways to apply machine learning, and there are many more out there in the literature. Have you come across a different application of machine learning in the petrophysics domain? If so, feel free to leave a comment on this article or get in touch on LinkedIn.

Thanks for reading!

If you have found this article useful, please feel free to check out my other articles looking at various aspects of Python and well log data. You can also find my code used in this article and others at GitHub.

If you want to get in touch you can find me on LinkedIn or at my website.

Interested in learning more about python and well log data or petrophysics? Follow me on Medium.

10 Applications of Machine Learning Within Petrophysics