Machine Learning May Solve the Cube Conundrum – Journal of Petroleum Technology

Optimal well spacing is the question. Well interactions are the problem. And cube drilling was supposed to be the answer. But it didnt turn out that way.

There was this idea that operators could avoid parent/child interactions by codeveloping their wells, said Ted Cross, a technical adviser with Novi Labs, during a recent presentation. They could develop many, many zones and maximize the recovery from a three-dimensional volume of rock.

This was cube drilling.

They could get a lot of operational efficiencies by having multiple frac crews on site, Cross said, building these megapads and saving on pad-construction costs.

The practice was tried, and, when the results were released, production was underwhelming. Stocks fell. Clearly, the cube was not the answer.

Nonetheless, much was learned from the venture into this dense drilling, which saw 50, 60, maybe 70 wells per section, within a given square mile, which is incredibly dense, Cross said. Just because the idea of a 70-well superdevelopment is dead doesnt mean that the concept cant still be useful.

While the concept of megapads has faded, it is not gone. Cross presented development maps and analysis that show people are still going to town on dense development, even if theyre not 60 wells per section. The industry has taken a little bit of time to figure out what geology supports these.

Consequently, well spacing remains important. Its still the key to driving net asset value and cash flow, said Novis president and cofounder, Jon Ludwig. If you go too aggressive, too many wells per section, obviously you lose cash flow, subtract net asset value, and, if youre public, you can subtract a good amount of company value as well. But, if youre not aggressive enough, you leave value on the table. So, its still critical to get this right.

Getting it right takes data, something the oil and gas industry has never lacked and something that cube drilling has produced in great quantities. Courtesy of all this cube development that has occurred, theres a lot of data, Ludwig said. Thats a huge advantage. We know now what good and bad look like. Every single cube thats been developed has left a signature.

Of course, the data doesnt help if it isnt used properly. We can all benefit from that if we know how to use the data well, Ludwig said. This is where machine learning comes in.

Machine learning models can tease out these subtle warnings from the past, Ludwig said.

One technique that benefits from the lessons of cube drilling is what Ludwig calls the surgical strike.

Getting cubes right is not all about a codeveloped cube in greenfield acreage, Ludwig said. A surgical strike, as weve defined it, is: What if I put a lease-line well between these existing developments? Or, what if Ive just acquired acreage in a very developed play like Eagle Ford or Bakken? How do I improve asset value? How do I bring learnings, completions designs, etc. how do I bring that in and actually improve net asset value by figuring out where you could still develop?

The machine-learning models help, Ludwig said, but the data must be dynamic. If youve built any kind of data-driven model, you want to use that model then to actually make forecasts and run scenarios for various ways you might develop your acreage. In order to do that, you need to have dynamic parent/child calculations for these hypothetical developments. If youre going to plan a cube where youre going to come in under an existing development, you need data that gets generated on the fly that describes distances, timing, etc. and allows whatever method youre using for modeling to change the forecast based on those factors.

This, Ludwig added, must be presented as a time series. We learned early on that making a point prediction is valuable and useful but its not nearly as useful as showing the shape of the curve and how the production rates change over time.

A cube, however, will not thrive in a black box. You really need to have the model not only output a forecast but also output something that explains why that forecast was made, what variables are driving that forecast, Ludwig said. He said that the models, if applied correctly, can explain their work.

What I mean by explain their work is: If a model forecasts X or Y, two different forms of a particular cube design, can it tell me also why? Because answering why is important when youre make the kinds of investment decisions that the industry is being asked to make. The sophistication of the models is not just the ability to make accurate forecasts, it is also the ability to explain their work. These two things together are critical for the financial case to continue to develop cubes.

See original here:
Machine Learning May Solve the Cube Conundrum - Journal of Petroleum Technology

Related Posts

Comments are closed.