3 ms·
Bluntly, they're useless except for a few niche cases. Anything they're capable of picking up _isn't_ what you're actually concerned about as an interpeter. S
by jofer 5y ago
Bluntly, they're useless except for a few niche cases.
Anything they're capable of picking up _isn't_ what you're actually concerned about as an interpeter. Sure they're good at picking reflectors in the shallow portion of the volume. No one cares about picking reflectors. That's not what you're doing as interpreter.
A good example is the faults in that video. Sure, it did a great job at picking the tiny-but-well-imaged and mostly irrelevant faults. Those are the sort of things you'd almost always ignore because they don't matter it detail for most applications.
The faults you care about are the ones that those methods more-or-less always fail to recognize. The significant faults are almost never imaged directly. Instead, they're inferred from deformed stratigraphy. It's definitely possible to automatically predict them using basic principles of structural geology, but it's exactly the type of thing that these sort of image-focused "automated interpretation" methods completely miss.
Simply put: These methods miss the point. They produce something that looks good, but isn't relevant to the problems that you're trying to solve. No one cares about the well-imaged portion that these methods do a good job with. They automate the part that took 0 time to begin with.
- deeviant 5y agoYou seem extremely biased against AI in general, to the point where I very much doubt anybody would benefit from hearing your opinions on it.
- jofer 5y agoI work in machine learning these days. I'm not biased against it -- it's literally my profession. I'm biased against a specific category of applications that are being heavily pushed by people who don't actually understand the problem they're purporting to solve. Put another way, the automated tools produce verifiably non-physical results nearly 100% of the time. The video there is a great example -- none of those faults could actually exist. They're close, but are all require violations of conservation of mass when compared to the horizons also picked by the model. Until "automated interpretation" tools start incorporating basic validation and physical constraints, they're just drawing lines. An interpretation is a _4D_ model. You _have_ to show how it developed through time -- it's part of the definition of "interpretation" and what distinguishes it from picking reflectors. I have strong opinions because I've spent decades working in this field on both sides. I've been an exploration geologist _and_ I've developed automated interpretation tools. I've also worked outside of the oil industry in the broader tech industry. I happen to think that structural geology is rather relevant to this problem. The law of conservation of mass still applies. You don't get to ignore it. All of these tools completely ignore it and product results that are physically impossible.
- jofer 5y agoIncidentally, I don't even mean to pick on that video specifically. I actually quite deeply respect the folks at Enthought. It's just that the equivalent functionality has been around and been being pushed for about 15 years now (albeit it enabled via different algorithms over time). The deeper problem is that it usually solves the wrong problem.
- toeget 5y agoI'm confused. You're saying they haven't incorporated basic physics into their model? How come? They don't have any geophysicists on staff?