You may have seen the news of the resounding win by the artificial intelligence program AlphaGo (built in the UK! Hooray!! Now owned by Google…..) over the South Korean world champion, the Go master, in the complex board game Go this month.

AI experts had predicted – last year, I think – that a computer program needed at least 10 more years of development before it would be able to beat a Go master. However, as I understand it, any rules-based system such as Go – and see below! – is a prime target for Machine Learning.

Hmmm, an interesting message there for the folks who have been telling me that it will take ‘a decade or more’ for AI to replace Subsurface Scientists in the oil & gas industry!

Below I offer my limited notion of Better Subsurface Science: how long before an AI system can do all of this?

Place your bets…….

Better Subsurface Science?

Ultimately, Subsurface Science can be summarised as ‘drilling profitable wells’

The key is to learn how to do this in a predictable, repeatable way (as opposed to drilling ‘on trend'; in a pattern; or effectively randomly).

We can classify the available methodologies as one of:

  • Rule-Driven Interpretation
  • Data Mining (using Analytics)
  • Modelling & Inversion

Rule-Driven Interpretation

Well-established ‘rules’ have been proven for Stratigraphy, Structural Geology, Sedimentology and describing Petroleum Systems (especially by creating GDE, CRS and CCRS maps). Nowadays these ‘rules’ are most commonly applied through seismic data, especially 3D seismic data.

The key ‘technologies’ are a) large quantities of inexpensive multi-client 3D seismic and b) commoditised interpretation workstations.

In truth, this methodology has now become completely commoditised: little commercial advantage accrues from getting it right, simply disadvantage flows from incompetent execution.Thus, if future competitive advantage is to be found, it must lie in either Data Mining – applying Analytics to data sets that are so large that they do not allow easy interpretation by humans – or Modelling & Inversion – especially those using and/or integrating more powerful geophysical technologies than towed streamer 3D seismic!

Data Mining

We can access satellite and airborne data, a significant variety of well results (logs, cuttings, core, flow rates), potential field, seismic, surface geology etc from a wide range of proprietary and public sources in diverse formats, with different accuracy, coordinate systems, & units of measurement.

We can be confronted with truly huge amounts of data and it is critical that we extract the critical information from all of it rather than looking at only a sub-set and/or simply entering the analysis with a ‘going-in model’ which we then look to authenticate.

Analytic techniques allow us to extract this critical information.

The key ‘technologies’ are a) the ability to integrate large quantities of diverse data and b) fast ‘Analytics’ applications, tuned to the problem in hand.

Modelling & Inversion

Predicting physical properties such as density, magnetic susceptibility, electrical conductivity, seismic velocity from geophysical data whether gravity, magnetic, electro-magnetic or seismic. Also addressing complex subsurface structures……

Of all these technologies, seismic remains the most powerful, offering the least ambiguity, the most resolution etc, and provides a framework into which other methodologies can be integrated.

The key ‘technologies’ are a) the integration of physically diverse multi-measurements and b) currently ‘niche’ inversion + modelling applications.

And so, we can make a conclusion about Who does this work, in an Integrated Team….

Rule-Driven Interpretation requires Seismic Interpreters and Geologists.

Data Mining requires Data Scientists.

Modelling & Inversion requires Geophysicists.

And this Team has to confront what I refer to as the Cloud of Points issue!

Thirty years ago, a ‘previous employer’ had an internal R&D project which rejoiced in the name of Lithology & Fluid Prediction (LFP).

Now LFP was founded on the idea that not only does seismic data show us geological geometries – folds, downlaps, onlaps, erosional truncation and the like – but that the very existence of reflections depends on rock physics, contrasts in impedances, and that we might get smart enough to predict actual lithologies and – wait for it – hydrocarbon content. And of course this notion has had some success, with AVO anomalies, flat spots etc etc.

However, I assert that we have not done as well as we might with such predictions and that this is primarily due to the relatively weak calibration that can be derived from well logs.

Who has worked in this arena and not found that well log-derived parameters such as sonic velocity, density or resistivity exhibit ‘cloud of points’ behaviour when plotted against for example depth? A ‘cloud of points’ through which it is a pretty brave person who fits a straight line or series of such lines, and then uses them to make lithology/fluid predictions?

Part of the problem has been selection and prejudgement. So many wells have penetrated, for example, the Kimmeridge Clay Formation in the UKCS that the problem only seems tractable if only a limited selection of them are used. And then a model is imposed – for example that the particular property will vary most strongly with depth or, perhaps, stress (if we have a way of calculating it).

Thus a sample of the available data is exposed to bi-variate analysis whereas the correct approach would be to subject all of it to a multi-variate analysis.

Maybe then we might even find some more oil in mature provinces such as the North Sea or South East Asia!

All in my humble opinion of course!

This article is for information and discussion purposes only and does not form a recommendation
to invest or otherwise. The value of an investment may fall. The investments referred to in this
article may not be suitable for all investors, and if in doubt, an investor should seek advice from
a qualified investment adviser. More

Source link