Machine Learning enriches Petrophysical Analysis: Five Case Studies

Machine Learning enriches Petrophysical Analysis: Five Case Studies

Machine learning (ML) is a method of data analysis that learns from data, identifies patterns, and makes predictions with minimal human intervention.?

The petrophysicist only needs to define the problem.?The computer guesses the answer and through successive iterations ‘evolves’ the best solution.?A ‘Fitness Function’ (FF) determines the difference between the ideal solution and its best guess, in a manner similar to how natural selection evolves the fittest species for a given habitat.?The ML minimises the FF and stops when it has found the best solution.

ML doesn’t require prior knowledge of the petrophysical response equations and is self-calibrating.?There are no parameters to pick or crossplots to make.?ML avoids the problem of “garbage in, garbage out” (GIGO) by ignoring noise and outliers.?ML programs work with an unlimited number of electrical logs, core, and drilling data.

Evolution of an Ideal Shaly Water Saturation Equation

A complex Middle East Carbonate field needed a bespoke shaly water saturation equation.?ML was used to ‘evolve’ an ideal equation.??

ML derives the form of the shaly water saturation equation and gives an independent estimate of SCAL parameters including the cementation exponent ‘m’ and the saturation exponent ‘n’.?In the presence of conductive shales, the resistivity measurement will be depressed. ?Shaly sand equations include a component related to shaliness to allow for this. ?The FF minimised the difference between core (red circles) and the ML derived water saturations. The ML evolved Shaly Sw Equation is shown by the blue line.

NMR Pattern Recognition

To successfully distinguish between oil and gas zones in real time, ML was used to reveal the fluid information hidden in the NMR T1 and T2 distributions.?

The Fitness Function for this application is “Determine the relationships between the NMR distributions and the core oil and gas saturations”, in a similar way to how human face recognition works. ML unlocked the relationships, shown in Tracks 5 to 7, to the core oil and water saturations, shown in Tracks 8 and 9 (green and red circles).? ML analysis of the NMR distributions predicted the oil and gas intervals as shown by the continuous curves in these tracks.

Shear Velocity Prediction

A North Sea field with 30 wells had shear velocity data (Vs) in only 4 wells.?Vs was required for reservoir modelling and well bore stability prediction.? ML was used to predict Vs in all 30 wells.??

The Fitness Function for this application is “Determine a relationship so that the predicted shear velocities are as close as possible to log derived shear velocities”.?The recorded Vs is shown in Track 4 in blue.?The predicted Vs is shown in green.?By incorporating high vertical resolution data, the Vs predictions were actually better than the recorded logs.?

Permeability Prediction

As it is not economically feasible to take core data on every well, ML was used to discover the relationships between logs, core, litho-facies and permeability.?As a consequence, field’s reservoir model was initialised with these all these data.

The Fitness Function for this application was ‘Determine a relationship so that the predicted permeabilities are as close as possible to core derived permeabilities’.?The ML first determined the litho-facies type as shown in Track 6, and then based on this result predicted the permeability shown as a continuous black curve in Track 4.?This compares favourably with the core derived permeability shown as green circles.?The prediction confidence is shown in Track 5, with the highest confidence shown in red and the lowest in blue.

The Quality Control (LQC) and Repair of Electrical Logs

Electrical logs can be automatically checked for errors and repaired using ML.

The LQC and repair of electrical logs is based on the premise that all logs are related.?ML is used to uncover these relationships, so that anomalies can be identified, and the correct log can be predicted.?The recorded density log, shown in green, is reading the mud density around X80 feet.?The predicted density log, shown in red, was derived by from all the other logs as if the density log was not run in this well.?Although the LQC is automatic the final step of replacing the poor log by the predicted log is currently left to the petrophysicist.

Full Paper

The Benefits and Dangers of using Artificial Intelligence in Petrophysics. Artificial Intelligence in Geosciences Volume 2, December 2021, Pages 1-10

https://www.sciencedirect.com/science/article/pii/S2666544121000125