DI Blog

Insights across the energy value chain

(Image credit: http://www.businesscomputingworld.co.uk)

Water saturation can make or break a play, but each petrophysical model may take your team hours or days to get right. How do you quickly get hundreds of control points to map water saturation across a basin with varying temperature, clay volume, and salinity?

Machine learning provides the toolkit needed

Neural networks are machine learning models that were inspired by the structure of our brains, improving with more data, bigger models, and longer computation times. At their core, neural networks test different ways of combining your data until they can best recreate a known solution. They are what power self-driving cars, automate cancer detection, and translate natural speech across different languages. These powerful algorithms have been proven effective across many disciplines and are well-suited to predict water saturation from standard well logs. In practice, we need to feed our neural network standard log suites along with petrophysically-calculated water saturation curves. The neural network can then learn how water saturation varies with changes to gamma ray, deep resistivity, and so on.

Creating the petrophysical models

Calculated petrophysical curves include Clay Volume, Total and Effective Porosity, Mineralogy, and Total and Effective Water Saturation.

Calculated petrophysical curves include Clay Volume, Total and Effective Porosity, Mineralogy, and Total and Effective Water Saturation.

Machine learning algorithms are useless without good data, so our first step was to generate up to 14 petrophysical models for each basin depending on the basin’s size. Each petrophysical model was made by first aligning the VClay curves calculated from Gamma Ray and the Density-Neutron crossplot. Guided by the M-N crossplot, we adjusted the clay points for sonic, density, and NPHI based on each individual well’s data. We built water saturation curves using the Poupon-Leveaux (Indonesia) equation and referenced the USGS produced water database for formation water salinity data.

Once our petrophysical models were calibrated to core data where possible, we trained the neural network to predict water saturation with several basic logs as the input data. However, the log suites themselves don’t contain information on spatial variations in temperature, salinity, clay density, and other variables that we set in our petrophysical models. We accounted for these geographic changes by simply including latitude and longitude of each well log in the model.

Model Accuracy

Model accuracy varied slightly since each basin had its own model with parameters optimized for that dataset, but the median absolute error was roughly 4% when predicting Total Water Saturation (SwT) and 7% when predicting Effective Water Saturation (SwE). Error likely increases for predicted SwE since there is a more complex petrophysical formula to calculate it. The most recent tests of these models show the coefficient of determination (R2) for predicted SwT above 0.92.

Predicted Sw curves and maps

Here’s a Transform cross-section of the Wolfcamp A in the Delaware Basin that shows SwT (pink) and SwE (blue) curves. Can you tell which three wells had their Sw curves generated by the neural network?

A mix of predicted and manual petrophysical SwT and SwE curves. The solution for which curves are predicted and which are manually-created is at the end of the article.*

A mix of predicted and manual petrophysical SwT and SwE curves. The solution for which curves are predicted and which are manually-created is at the end of the article.*

The detailed section below shows the variance between SwT (pink, predicted total water saturation) and SwE (blue, predicted effective water saturation).

Using just the manually-created Sw curves gives very little detail throughout the basin (below, left). With over 200 additional predicted Sw curves (below, right), we can get a much more complete understanding of how water saturation varies throughout several of the major basins in the US.

Part of an SwE map containing 14 manually-created curves (left), and the same area from a map containing 225 additional predicted SwE curves (right).

Part of an SwE map containing 14 manually-created curves (left), and the same area from a map containing 225 additional predicted SwE curves (right).

In Transform, we created water saturation maps by averaging both the manually-created and predicted Sw curves over zones of interest where DPHI was above a cutoff set for each basin. The predicted logs gave us much-needed control for average water saturation and net feet of pay maps. Together these two maps allow us to define the play boundaries and can be used for a quick assessment of field development viability. As of today, full-sized maps and all the data needed to recreate them are available for the Delaware Basin, Central Basin Platform, and Mid-Continent Play Assessments. Future releases will add coverage in more basins, stratigraphic intervals, and manually-created petrophysical logs, as well as tweaks to these models to better predict water saturation.

*The three middle wells are displaying predicted Sw curves.

The following two tabs change content below.