b'Seismic window Seismic windowand I could produce fault maps at thethis led to artefacts that obscured the click of a mouse button (Figure 1). major trends (e.g. thinned fault likelihood).Over the last decade there has been aThe ML implementation creates a great deal of movement in designingseismic volume of the probability of a methods to highlight faults and fractures.fault at a certain location. It produces a Originally a coherence volume wasclear fault/no fault definition with very calculated, but I believe Amoco held thefew artefacts (Figure 2). How it does this copyright for the method and so otherI wouldnt know; Ill just take the results companies had to find alternatives suchfor granted.as similarity. At the time the similarity volume was providing some excellentActually, I do know a bit. The machine results, especially when viewing timelearning solution is a convolutional Michael Micenkoslices or horizon slices. But the usefulnessneural network algorithm that I believe Associate Editor for Petroleumof similarity, semblance and coherencyhas been trained on a database of faults micenko@bigpond.com cubes was often compromised by noisyworldwide. Regardless of how it works, data that resulted in variable quality of aIm impressed that it worked straight fault trace vertically and laterally. Someout of the box with no input required innovative techniques were sometimesfrom me. I really am being replaced by Artificial intelligenceused to sharpen up the fault image, butamachine.finds faultsIn recent years it has taken some effort for me to get my head around Machine Learning (ML). I even bought a book (The Master Algorithm by Pedro Domingos) that was recommended to me by a Canadian mate. He said it made understanding ML simple. I found it a great cure for insomnia, and after a few weeks reading I came to the conclusion that nothing will make ML simple and it was a book for long cold Canadian winters. But then, a few months ago, I was fortunate enough to try out some ML software that supposedly could highlight fault planes in a seismic volume. It was quick, worked a treat, and perhaps the best thing about it was that I didnt have to know anything about the maths.Figure 1.Display of ML produced faults in vertical plane and similarity data in horizontal plane. The Suddenly I had become a Nintendo geosimilarity derived faults are visible but fuzzy and obscured by noise.Figure 2.Comparison of various methods of detecting faults. Clockwise from top left a) Similarity, b) fault likelihood, c) thinned fault likelihood and d) Deep learning artificial intelligence solution.29 PREVIEW AUGUST 2020'