Oil IT Journal interview—Tom Smith, GeoInsights

Founder of SMT and fledgling Geophysical Insights expounds on new pet topic—neural networks.

What have you been up to since selling Seismic Micro Technology (SMT)?

I stayed on for a few months to assist with the reorganization, I’m still on the SMT board. After that I began doing what I enjoy most—solving geophysical problems. The late Tury Taner of Rock Solid Images was inspirational and he helped me understand the theory and practice of neural net techniques—that was a high point in my professional career.

Geophysical Insights (GI) is separate from SMT?

Absolutely. GI is involved in geophysical research. We will be presenting a paper at next month’s SEG.

What are your other R&D areas of interest?

I can’t talk about this yet but we are working on challenges and on truly fundamental geophysical research questions. We are not taking the university/JIP route. Our clients are looking for a vehicle for research collaboration on such issues. We are also working with Sven Treitel on some pretty deep topics.

Is GI’s neural net technology pre or post stack or what?

It is essentially multi-attribute work. People can easily generate 20, 30 or 50 attribute volumes so for any time (or depth) you have a multiplicity of spatially distributed attribute values. Now I defy anyone to interpret more than 2-3 attributes at a time! You can do flicker type comparisons—but these remain a challenge to interpret. Interpreters like to leave no stone unturned—and need help from something like a neural net to do, say, spectral decomposition, with up to 20 frequency bands—and then to ‘ground truth’ the results against well logs. We like to think of seismics as a cylinder of data surrounding a well bore—maybe with 30 or so attributes. These need to be matched with another set of ‘attributes’ coming from borehole data—extending cross plots to many dimensions.

We have often wondered just how many truly independent (and meaningful) attributes can be computed from the limited number of field measurements in the seismic trace…

This is the ‘saliency’ issue in a neural nets and in machine learning. How much redundancy do you need and what makes up truly independent attributes. One tool to investigate this is principle component analysis. This has been around for a while but has had limited application. It is easy to run the wheels off the trolley! We use an unsupervised neural net and look for natural clusters in the data. The answer is, it all depends, some attributes are significant some of the time. Elsewhere, in other geographies, they may not work so well.

Any flagship client/projects?

Not that we can talk about yet. But Tury was working on this for a decade—using neural networks for lithology identification.

Basically, neural networking boils down to data mining…

Exactly. Supervised neural nets have been used in optical character recognition and in bibliographic text mining through thousands of scientific articles—using key word counts and clustering in terms of relevance into topic maps. It is easy to see how neural nets can be used to extend attribute analysis across a company’s 3D data resource—using it on all 3D surveys to understand different areas of the world and build a knowledge base. Neural net techniques can also be across a well log repository to check for a particular response and build company-wide wisdom. Note that there has been some reticence to neural nets and artificial intelligence. But such techniques are not intended to ‘replace’ the interpreter. We just want to get past the drudgery of interpretation—rather like when we went beyond manual digitization to the workstation. Neural nets can also be used for business trends in data. Wal-Mart is said to have a larger database than the CIA! More from www.geoinsights.com.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.