Why We Need Machine Learning in Oil and Gas Industry?

The Process of Correlation Generation:

The Era Pre-Machine learning and AI in oil and gas industry mainly involved cumbersome research and correlation derivation from the experimental work, and then those correlations if they were good enough, they would find their way into technical papers for other engineers to use in oil and gas industry, example of those would be:

Reservoir Engineering:

1. Arps Decline Curve Equation
2. Shale Volume Equations

Production Engineering :

1. Fluid Flow Correlations(Beggs&Brill, Aziz&Govier)
2. Inflow Performance IPR(e.g Vogel, Fetkovitch, BackPressure)
3. Critical Velocity in Liquid Loading (Turner, Coleman, Dugan)

Those methods are still in use in oil and gas industry because they are proven, and they provide good results (assuming good data quality).However, these methods have been created at very restricted test/experiment conditions (Cost is the main factor in this limitation). methods have been created at very restricted test/experiment conditions (Cost is the main factor in this limitation).

One thing to keep in mind those correlations have stop evolving and they are provided as is. Now the only thing we can do with these in to validate them and tweak them. Unlike Machine Learning the Model is recalculated upon new data arrival.

The New Challenges

New challenges in oil and gas industry are the volume of data that are being generated from oil and gas fields, the utilization of these data as is (without aggregation) is somewhat challenging in oil and gas industry, since the oil and gas correlations are designed in such a way it's really easy to use.

Knowing that, this means in order to use Arps DCA we need to convert the production data from DOF sensors into Daily or Monthly aggregation to get that smooth curve we love. But this is not always a good practice. (Take a look at the image bellow)

Other Challenges would include shale production, CBM and other unconventional resources that usually requires a certain set of empirical equation that just works on that. This is time consuming and it's an art and an engineering design to create such correlations.

We would love to include all of our Realtime data into a predictive model that can predict the near future parameters for us and this is where machine learning can shine.

The need for Machine Learning

Take a look at the unstable production bellow, there is no equation in oil and gas literature that can predict this kind of behavior (Sure thing you can use commercial simulator for that)

Machine learning can be a help in this situation, because machine learning is data aware meaning that all the data that you are considering will be used for the training and test of these model, plus some machine learning models are in the state of aware of the previous production periods before making prediction for the newly forecasted points.

These models can use detailed data with all the performance parameters(features) the engineer wishes to use and these can be :

1. Water Production
2. GOR
3. Yield Oil Gravity in the Facility
4. Water Salinity.
5. and many more.

An Example of Labeled Data: Classification Problem

In 1969, Turner published his Infamous paper Read Here on OnePetro and that paper revolutionized the way we are dealing with liquid loading analytics, he proposed a very simple equation(correlation) that can be used to predict liquid loading issue. Take a look at the figure bellow:

The plot above was the basis of his equation, this was a remarkable achievement back in the 60's where access to computers and analytics where very limited, but do we actually get correct results from this? Please Check this Liquid Loading By Turner Calculation.

Turner's equation was based on some ~110 data points (How i know this? I Digitized the data from the original paper).Nowadays we are making monetary investments and really important decisions based on these 110 data points.

Now days we get thousand upon thousands of data points each week, we can convert those into actual classification problem that we can utilize to classify any gas or High GOR wells base on their loading status.

We can generate decision maps and well labels based on the large volume of data we are getting now days(the more data you have the more understanding and better your ML model will be).

Addendum

It's good to use Machine learning for complex problems where there are 3,4,5 or more variables that are involved in the analytics, plus machine learning is fast and you can get results within seconds. Since ML Models are complex in their nature(some of it at least), we cannot get equations back from these, but we can export the model to a file and use it anywhere we want(as long as we have python in that machine).

It's not recommended to use ML for each and every task(e.g calculating OOIP with primitive data volume).

Machine Learning and Industry Correlations can go hand in hand when used through Physics informed Models PIM.