Surely you've seen the improvement over the last 5-6 years from machine learning in all the interpretation toolsets. The last place I worked we internally had a seismic inversion tool that blew all the commercial suites out the water. I'm currently contracting for an AI/ML service company currently that has a synthetic welllog tool that is can apparently beat the pants off actual well logging tools for a fraction of the cost (though I'm not a geologist or petrophysicist so I can't personally verify this.)
I think the problem is more the media and advertisers likes to paint the picture of a magical AI tool which will instantly solve all your problems and do all the work instead of a fulcrum to make doing the actual work significantly easier.
Bluntly, no. There hasn't been an improvement. At all.
We've been using machine learning in geology for far longer than it's been called that. Hell, we invented half the damn methods (seriously). Inverse theory is nothing new. Gaussian processes have been standard for 60 years. Markov models for stratigraphic sequences are commonly applied but again, have been for decades.
What hasn't changed at all is interpretation. Seimsic inversion is _very_ different from interpretation. Sure, we can run larger inverse problems, so seismic inversion has definitely improved, but that has no relationship at all to interpretation.
Put another way, to do seismic inversion you have to already have both the interpretation _and_ ground truth (i.e. the well and a model of the subsurface). At that point, you're in a data rich environment. It's a very different ball game than trying to actually develop the initial model of the subsurface with limited seismic data (usually 2d) and lots of "messier" regional datasets (e.g. gravity and magnetics).
I am wondering (knowing nothing about this) if there is an issue with the approach to acquire data that it putting AI in a difficult position. This is akin to trying to train and AI to walk in the footsteps of a geophysicist, rather than making new footsteps for the AI. I guess I would extend this to radiology too since it seems to be the same issue.
Let me give an example:
People often mention that truck drivers are safe from automation because lots of last mile work is awkward and non-standard, requiring humans to navigate the bizarre atypical situations the truck encounter. Training an AI to handle all this is far harder than getting it to drive on a highway.
What is often left out though is the idea that the infrastructure can/will change to accommodate the short comings of AI. This could look like warehouses having a "conductor" on staff who commandeers trucks for the tricky last bit of getting on the dock. Or perhaps preset radar and laser path guidance for the tight spots. I'd imagine most large volume shippers would build entire new warehouses just to accommodate automated trucks.
A long time ago people noted that horses offered much more versatility than cars since roads were rocky and muddy. How do you make a car than can traverse the terrain a horse does? You don't, you pave all the roads.
Automatic interpretation has been a thing for decades and the promise of replacing a geoscientist completely is always just over the horizon. Even with DL. The new tools are better yes, but honestly I wouldn’t invest in this space. Conventional interpretation is dead in the US. All the geos got laid off.
I’m going to call bullshit. No artificially generated well log is going to ever be better than a physically measured log.
I agreed with you up until the last paragraph. Generating data that cannot be told apart from any real data, cave ins and all, is probably one area where this can succeed.
Your other comment about picking the horizon of interest is really on point, that's where automated interpretation as it's been buzzworded to hell to date as had had no chance and has never lived up to how it was pitched. Many tools just made the problem worse.
That fact that this might change in some distant future, well it may be solved in some capacity but is it really worth the effort? Given that exploration has a limited future.
I think the problem is more the media and advertisers likes to paint the picture of a magical AI tool which will instantly solve all your problems and do all the work instead of a fulcrum to make doing the actual work significantly easier.