At some point should write up thoughts on how sedimentology informs basic architecture and feature creation

[JustinGOSSES]

Write up a paragraph(s) about how the spatial and temporal variation in allogenic and autogenic sedimentary controls and the number of wells available for training makes it unlikely straight pattern matching, curve matching, or more mathematical approaches will work for stratigraphic correlation. Additionally, these same discussion points can point towards what sort of features should be included and what type of model to build.

Include points about:

1. Curves can easily be dominated by sandstone and other lithologies that don't extend very far and give you short spatial signals that are more likely to be locally driven and not reflective of spatially extensive time uniform forcing. What a geologist typically does is ignore large parts of the curve and look at the shale parts that are more likely to reflect basin-wide and temporal signals. Our methods need to include this somehow.
2. Allogenic signals tend to be basin-wide but may not be felt at the same time across all parts of the basin or expressed in the same way. Ideally, our methods need to include this idea.
3. Tops have some characteristics that are ideally basin-wide, some characteristics that are different in different sub-regions, other characteristics that gradually change over a distance, and other changes that are somewhat or completely randomly spatially distributed. How do we create a system that can identify for itself which types of characteristics or comparisons of characteristics are in each of those groups and make predictions for a well it hasn't seen?
4. This type of technique can't be done without tops already existing in at least several hundred if not thousand of wells. It also requires the wells to penetrate the same formation with a relatively similar depositional environment in order to build up patterns. Some characteristics or patterns will be rare in the dataset. How do we make a system that knows when to fail gracefully and use something very basic like the average thickness of neighbors rather than latching onto a similar looking curve bit hundreds of meters farther down?
5. Although we are trying to do chronostratigraphic correlation and not lithologic correlation. Are there facies prediction techniques that can be used here. For example, is the shale different between McMurray and Wabiskaw big enough to recognized as separate shale facies and then use that as a feature?

[JustinGOSSES]

some early thoughts on this from an earlier blog-post http://justingosses.com/thoughts-on-machine-learning-predictions-of-chronostratigraphic-surfaces/ This could be cleaned up, expanded, and more directly used to explain why other methods fail in some circumstances and not others.