https://library.seg.org/toc/leedff/38/7 In case access is an issue, here is a quick takeaway (excerpt from The Leading Edge, July 2019)

1) demystify machine learning (ML) - using wedge model to illustrate thin-bed tuning effect, when the underlying physics is well understood then conventional inverse methods generally lead to a better solution, but when the physical model does not adequately describe the real world, or the inverse problem is nonlinear, then machine learning could succeed where traditional methods fail.

2) frac hit identification - rather than just use the raw data, when employ domain knowledge to engineer features that highlight the fracture hits and score well on the test data. Even for a relatively small data set, a ML model can be successfully trained to detect the events, a process that is time-consuming and highly subjective when done manually.

3) seismic fault interpretation - train their convolutional neural network (CNN) using only synthetic seismic data. Training with synthetics avoids the time and expense of acquiring appropriate, labeled field data for training the network. The results are sensitive to the parameterization of the wavelet used in training, but that the CNN can predict even the low-frequency information that is missing in field data.

4) transfer learning - compare and contrast the application of deep learning with traditional forward and inverse problems in geophysics. The two approaches share some numerical methods, objective functions, and regularization terms. Similar deep learning network architectures can be applied to different tasks in seismic interpretation, namely automatic horizon picking and interpolation of lithology between wells using seismic data.