Live from Improving Modeling and Predicting Reservoir Behavior -

dataVedik hands-on ML exercise highlights -

1. Setup - 6 wells with (lat, long) and gamma ray values sampled in depth (GR, depth) TD ~10,000 ft.
2. ML - use Random Forest Regression, train on (depth, lat, long); predict GR at each sample
3. Get reasonable result.

While this is a simplistic ML exercise, it illustrates two things:

a) interpolation - as in this exercise, think spatially restricted / "skinny" model, applicable to pancake geology. Won't work too far away from the training well. (Otherwise, flatten data / remove structure in data prep.)

b) extrapolation - try spatially "loose" model, may opt to train on only (lat, long), and predict GR values to restricted depth range (e.g., some 1,000 ft interval). With such understanding, ML may get better prediction when stepping further away from the training well.

That is just one interpretation. Feel free to chime in.